reasampler/CONTEXT.md

ReaSampler — implementation briefing

A native C++ REAPER extension that captures any arbitrary audio source into a per-project sample bank (cached files + a docked grid), decoupled from the arrange view, with keyboard/MIDI-bindable capture and placement. Built as a precision tool: deterministic, non-destructive, no clutter.

This document is the spec. Read the existing scaffold first (src/main.cpp is the REAPER<->extension contract; the pure/testable-core split in src/mpe_model.* is the pattern to preserve — the MPE model is being replaced, the discipline is not). Verify every REAPER API name and signature against vendor/reaper-sdk/sdk/reaper_plugin_functions.h before use — the API names in this brief are correct-by-intent but treat them as hints, not gospel, and check argument order/types.

The load-bearing principle

Capture and placement are separate acts. Capturing audio writes a file to the bank and adds an index entry — it NEVER puts an item in the arrange. Placement is a distinct, on-demand action (insert / drag). Any code path that auto-inserts a capture into the timeline violates the entire point of the tool and must be rejected in review. This single rule is why the tool exists.

Settled decisions

• Capture modes: offline render (deterministic, the default) AND realtime record (for hardware / performed FX). Both sit behind one capture interface and produce identical bank entries.
• Bank scope: per-project, travels with the .rpp. Files live in a project-relative subfolder; the index persists in project ext state. No absolute paths anywhere in the index.
• Material: must handle full-mix/stem bounces, chops/one-shots, and single-cycle/wavetable grabs equally. That means exact sample-accurate bounds, explicit tail control, channel-count preservation, and loop/zero-crossing handling all matter from day one.

Module architecture

Preserve the scaffold's split: pure, REAPER-free logic in one set of files (unit-tested outside the DAW via the existing tests/ + CTest harness), REAPER- facing shells in another.

Pure (no REAPER types, fully unit-tested):

• bank_model — the Sample metadata struct and the BankIndex (add / remove / query / tier moves / dedup-by-hash) plus JSON serialize/deserialize to a std::string. This is the heart; test it hard.
• peaks — compute waveform min/max bins from raw PCM. Feed it a known signal (sine, ramp) and assert the envelope. We compute our own thumbnails from the captured file rather than depending on REAPER's peak API — we own the file format, so this is simpler, testable, and dependency-free.

REAPER-facing:

• capture — the ICaptureBackend interface plus OfflineRenderBackend and RealtimeRecordBackend. Input: a CaptureRequest (capture scope — item or track, time range, tail, SR/bit-depth/channels, output path). Output: a finished file + a populated Sample handed to bank_model. Capture is always wet; the FX scope (not a wet/dry dial) is the control — the pure render_settings module maps a scope to its render-source bits and its FX-bypass plan (see §Capture FX scope below).
• insert — placement via InsertMedia; conform-to-project-tempo vs literal, as an explicit flag (never silent stretching).
• bank_panel — the docked grid: LICE-drawn thumbnails, audition, multi-select, keyboard navigation. Reuses the docking setup already in mpe_view.cpp.
• persist — project ext state <-> bank_model JSON; project-relative path resolution (resolve bank folder from the current project path).
• actions — registers the capture/placement/slot action family and routes each to the modules above (the command_id + gaccel + hookcommand pattern from main.cpp).

Data model (sketch — refine in code)

Sample: id, display name, relative file path, source mode, source range (start/ end in project time + PPQ), track GUID(s) if applicable, wet/dry, channel count, sample rate, length (seconds + musical/beats), capture tempo, optional key, peak/RMS/LUFS, clip flag, tier (scratch | archive), content hash, provenance (parent sample id + FX-chain snapshot string, when resampled from another sample), created timestamp.

BankIndex: ordered collection of Sample, keyed by id; hash lookup for dedup; tier filtering; JSON round-trip. Scratch tier is auto-prunable; archive is kept.

REAPER API surface (verify all signatures)

Offline render (the crux — prototype this first):

• Drive render settings with GetSetProjectInfo (RENDER_BOUNDSFLAG, RENDER_STARTPOS, RENDER_ENDPOS, RENDER_TAILFLAG, RENDER_TAILMS, RENDER_SRATE, RENDER_CHANNELS, RENDER_SETTINGS) and GetSetProjectInfo_String (RENDER_FILE, RENDER_PATTERN, RENDER_FORMAT). The RENDER_SETTINGS bit choice is driven by capture scope: item scope renders selected media items (single-file), track scope renders selected tracks via master. There is no master-mix scope (to capture the master, render a track). Render is always wet — REAPER has no true pre-FX "dry" render bit; FX scoping is done by the FX-bypass-around-render mechanism, not a render bit (see §Capture FX scope).
• Trigger a no-dialog render via the appropriate render action / RENDER_SETTINGS bit. Confirm the exact command id and the "render without opening dialog" flag against current REAPER — do not assume; test that it runs headless.
• Determinism is a hard requirement: two identical requests must produce bit-identical files (enables the null test below).

Realtime record (track scope — item realtime is deferred):

• Taps the selected track, not the master. Recipe: create a hidden temp track, route a post-fader send from each selected source track into it (CreateTrackSend, default post-fader/full-stereo — captures each track's own output before the parent/master sums it, so the tap is chain-independent by construction and needs no FX bypass), set the temp's record mode to record-output (latency-compensated, B_MAINSEND=0 so it does not sum back), arm (I_RECARM), CSurf_OnRecord, run for the range, OnStopButtonEx, then move the recorded file into the bank and delete the temp track (which drops the sends — no source track is left mutated). Verify I_RECMODE values for output-recording.
• Master→track sends are refused by REAPER as feedback loops — this is why the tap is the selected track's post-fader output, not the master.

Sources & metadata:

• Time selection: GetSet_LoopTimeRange. Razor edits: GetSetMediaTrackInfo_String(track, "P_RAZOREDITS", ...). Tempo: Master_GetTempo / TimeMap2_timeToBeats / GetProjectTimeSignature2. Selected items/tracks: CountSelectedMediaItems / GetSelectedTrack.

Placement:

• InsertMedia(path, mode) at edit cursor / new track / replace selection (verify mode bits). SetEditCurPos. Wrap edits in Undo_BeginBlock2 / Undo_EndBlock2.

Persistence & paths:

• SetProjExtState / GetProjExtState (namespace e.g. "reasampler") for the index JSON. Resolve project folder via EnumProjects / GetProjectPathEx; store the bank under a project-relative subfolder; keep only relative paths in the index.

Build order (each milestone independently testable)

1. bank_model + JSON round-trip + unit tests. (pure — no REAPER)
2. peaks + unit tests. (pure)
3. Offline capture of the time-selection master mix to a wav in the project bank folder; add a Sample; log it to the console. (the render-driving spike)
4. persist: write the index to proj ext state, reload on project open; confirm it survives Save / Save As. (bank travels with the .rpp)
5. bank_panel: docked grid with thumbnails, audition, selection.
6. insert: "insert selected sample at edit cursor" action via InsertMedia.
7. Capture action family: two FX scopes — item (item/take FX only) and track (item FX + the selected track's own track FX), each chain-independent (the out-of-scope chain is neutralized to unity for the render). No master scope — to capture the master, render a track. Each scope captures over a time-or-razor range (razor is a range source, not a mode: razor-when-present, else time selection). Capture is always wet; the scope is the control. All registered as bindable actions (tail off; a tail-on variant is a later opt-in).
8. RealtimeRecordBackend behind the same interface.
9. Slots: "capture to slot N" / "insert slot N", MIDI-bindable (MPC-style).
10. Provenance + "re-capture from source"; null-test verify action.
11. Polish: batch capture (per selected item / per razor area), resample-and-mute- source, conform-on-insert, drag-out to OS.

Precision invariants (enforce, test)

• Null test: a dry offline capture of a range, re-inserted at its source position, nulls to silence against the source. Ship this as a verification action; it is the tool's trust anchor. (Verification action cut per docs/product/provenance.md — manual verification only.)
• Bit-identical repeats: identical offline requests produce identical files.
• Non-destructive: capture never mutates source items or tracks (realtime's temp track is created and removed cleanly; source routing is restored).
• Exact bounds: no rounding of the requested range; no added silence unless a tail is explicitly requested; channel count preserved (no silent stereo fold).
• Relative paths only in the persisted index.

Non-goals / guardrails

• No auto-insertion of captures into the arrange (see the load-bearing principle).
• Native OS drag-out is deferred to the final milestone — InsertMedia-driven placement is the primary path and must work first.
• Do not depend on REAPER's peak API for thumbnails; compute from the captured file.
• Do not silently time-stretch on insert; conform is opt-in.
• Do not trust this brief's API names blindly — verify against the SDK header.

Open questions to resolve during build

• Exact no-dialog render command/flag on the current REAPER build.
• Realtime record routing that captures wet master output without altering the user's monitoring. Resolved: realtime taps the selected track (track scope), not the master — a post-fader track→temp send is chain-independent by construction, so there is no wet-master-routing problem and monitoring is untouched (see the Realtime record API surface above).
• Audio format for the bank (wav bit depth default; allow float for wavetable fidelity).
• Thumbnail cache: recompute vs store peak bins alongside the index.

---

Design View — additive phase spec

Additive section. This is a standalone phase parallel to — not part of — the M0–M11 capture roadmap above. Nothing above changes. Product framing (workflow narrative, screenset differentiation, N-mode reasoning, design-direction calls) lives in docs/product/design-view.md; this section is the authoritative technical spec, matching the house style of the capture spec. Same standing discipline applies: verify every REAPER API name/signature against vendor/reaper-sdk/sdk/reaper_plugin_functions.h before use.

What it is

A track-visibility-plus-processing "mode" system, toggled from the ReaSampler window. Tracks used purely for sound design (scratch oscillators, FX mangling, resampling sources) are tagged into Design mode; the arrangement's real tracks are Arrange mode (the default). Toggling to a mode hides and disables the tracks that don't belong to it. Workflow value: mental separation + clutter elimination — be in Design view, resample into the bank, flip to Arrange, place it.

Design View is a mode projection over REAPER's single arrange timeline — reaching both tracks (parked per mode) and items (lane-split per mode on shared tracks). It approximates two canvases without a second surface: same project, same timeline, but each stance sees its own tracks and its own items. It never duplicates the project or opens a second window.

Two-canvas reach — settled (2026-07-23). The original single-canvas framing here was derived from a REAPER constraint, not chosen as a product stance. Daniel reopened it — "as close as possible to two separate design and arrange canvases, same project, different items and leaves" — and has now signed off on the mechanism and resolved all forks. The "different leaves" half is delivered by track-parking (D1). The "different items" half is delivered by per-item mode membership via REAPER fixed lanes (track-side I_FREEMODE=2, I_NUMFIXEDLANES, C_LANEPLAYS:N; item-side I_FIXEDLANE, C_LANEPLAYS, B_FIXEDLANE_HIDDEN): each mode owns a lane on a shared track, and toggling shows/plays only the active mode's lane. This is an item-visibility projection over the one timeline — the exact analog of today's track-visibility projection — with still no literal second surface, window, or duplicate project. It is an additive sub-phase (Phase D2 / Phase E) on top of D1, spec'd in §Two-canvas sub-phase below. Full framing and the rejected alternatives (timebase-offset, subproject) live in docs/product/design-view.md §Two-canvas direction.

It is the visibility/processing analog of the capture pillar's load-bearing rule: designing and arranging are separate stances on one timeline, and the tool enforces the separation without ever destroying the user's real state.

Settled decisions

• Membership. Default = Arrange; every untagged leaf belongs to it. Leaves opt in to Design (or any mode). No track appears in two modes at once except (a) via an explicit show-both toggle, or (b) parent/folder derivation.
• Parents are derived, never tagged. A parent/folder track is visible in mode M if either (a) any descendant leaf is visible in M (descendant-derived), or (b) the parent belongs to M by its own membership — and an untagged parent is an Arrange member by default. So an untagged folder carrying its own FX/media above all-Design leaves shows in both Arrange (its own default) and Design (derived from its children). A parent is never parked in any mode it is visible in. Rule of thumb: tag leaves; parents follow — the common case, since a content-bearing folder still surfaces wherever its own membership places it. Master track is always visible and never touched.
• N-mode model, two-mode UI. The data model carries arbitrarily many modes; the UI ships Arrange + Design. A mode is (stable id, display name, ordinal). Arrange is special only as the default home for untagged leaves.
• Parking a track (inactive-mode leaf): B_SHOWINTCP=0, B_SHOWINMIXER=0 (hide both panels), B_MAINSEND=0 (out of mix), I_FXEN=0 (FX bypassed), and TrackFX_SetOffline(track, fx, true) for each FX (reclaim CPU). Full CPU-park is the deliberate choice over mix-removal-only.
• Never touches B_MUTE / I_SOLO. The tool owns only visibility, B_MAINSEND, I_FXEN, and per-FX offline — on every managed leaf, tagged or untagged. User mute/solo survives every toggle untouched.
• Persistence. Membership index + last-active mode + per-track flag snapshots ride in the existing "reasampler" project ext-state namespace and travel with the .rpp. On project open, reapply the active mode's visibility + processing.

Precision invariants (enforce, test)

• Non-destructive restore. For every flag the tool drives, snapshot the prior value before parking; on toggle-back restore from the snapshot, never to a hardcoded "on." Round-trip (snapshot → park → restore) returns every driven flag to its captured value. This is the phase's trust anchor — the analog of the capture null test — and is enforced in the pure layer.
• Mute/solo untouched. No toggle ever reads or writes B_MUTE / I_SOLO.
• Master untouched. The tool never drives the master track's visibility flags (the SDK forbids B_SHOWINTCP/B_SHOWINMIXER on master; the invariant agrees).
• GUID-keyed, reorder-safe. Membership keys on track GUID (GetTrackGUID), never track index; tolerates unknown/stale GUIDs (prune on reconcile).
• Relative/portable state only in the persisted view section (GUID strings, mode ids — no absolute paths, no index positions).

Documented caveat

Offlined FX re-instantiate when a track returns to the active mode. Stateful plugins (convolution, loaded samplers, tail-holding effects) re-initialize on return — possible load hitch, un-persisted internal state lost. Accepted cost of the CPU reclaim; surface it at the toggle affordance (tooltip).

Module architecture (preserve the pure/shell split)

Pure (no REAPER types, unit-tested — the mirror of bank_model):

• view_mode_model — mode registry (id/name/ordinal; Arrange + Design seeded); membership index (track GUID → { mode ids } + per-track show-both flag; add / remove / retag / query); folder-tree-aware visibility derivation (given the current parent↔child tree supplied by the shell + the active mode, compute the visible set); the parking/restore planner (given active mode + snapshot record, emit the exact (track, flag, value) operation lists for park and restore — where the restore invariant is enforced); JSON round-trip of modes + membership + show-both + snapshots + active mode.

REAPER-facing:

• view shell — reads I_FOLDERDEPTH across the track list to build the parent↔child tree and feeds it to view_mode_model; applies the planner's operations via SetMediaTrackInfo_Value (B_SHOWINTCP / B_SHOWINMIXER / B_MAINSEND / I_FXEN) and TrackFX_GetCount + per-FX TrackFX_SetOffline; snapshots prior flag values before parking; resolves GUIDs via GetTrackGUID / guidToString / stringToGuid. Never touches master visibility, never touches B_MUTE / I_SOLO.
• persist (slice) — serialize/deserialize the view section into the "reasampler" namespace alongside the bank; on project open, rebuild the tree and reapply the active mode.
• actions (entries) — toggle active mode; activate mode: Arrange / Design; tag/untag selected tracks → mode; show-both for selected tracks. Registered with the command_id / gaccel / hookcommand pattern; toggle + mode-jumps MIDI-bindable.
• UI (in the ReaSampler / bank_panel window) — a segmented mode switch ([ Arrange | Design ]) in the window header, active segment lit; small per-mode membership count; the offlined-FX caveat as a tooltip. Tag/untag acts on the current REAPER track selection, not a per-track widget.

Show-both semantics

A per-track "pin visible across modes" flag that re-enables processing whenever shown. A show-both leaf appears in every mode's visible set and is never parked — its driven flags stay at snapshot/restored values, FX online, in the mix. ("Show but keep parked" is not offered — a visible-but-silent-and-offline track is clutter with a thumbnail.) Stored on the membership record; persists; togglable per selection.

REAPER API surface (verify all signatures)

• Visibility/routing/FX flags via GetMediaTrackInfo_Value (snapshot) / SetMediaTrackInfo_Value (apply): B_SHOWINTCP, B_SHOWINMIXER, B_MAINSEND, I_FXEN. (Note: brief cited B_SHOWINMCP; verified SDK name is B_SHOWINMIXER.)
• Per-FX offline: TrackFX_GetCount + TrackFX_SetOffline(track, fx, offline).
• Folder tree: read I_FOLDERDEPTH per track to derive parent↔child structure.
• GUID keying: GetTrackGUID, guidToString, stringToGuid.
• Persistence: SetProjExtState / GetProjExtState under "reasampler" (shared with the bank index — one blob, two logical sections).
• Wrap flag mutations in Undo_BeginBlock2 / Undo_EndBlock2 as appropriate.

Non-goals / guardrails

• No literal second canvas. A literal second arrange surface, a second window, or a duplicated project stays rejected — reject any such path in review. This guardrail was narrowed (2026-07-23), not lifted: per-item mode separation via REAPER fixed lanes on shared tracks (an item-visibility projection over the one timeline) is now allowed and specified in §Two-canvas sub-phase below. Paths that remain rejected: subproject / second-project-file approaches, and overloading item D_POSITION with mode semantics (timebase-offset regions) — the latter collides with the capture null test. See docs/product/design-view.md §Two-canvas direction for why those were rejected.
• Never touch mute/solo. Any code path reading/writing B_MUTE / I_SOLO is a bug.
• Never drive a manual lane (Phase D2). A mode toggle touches only managed lanes (minted by the mode system, keyed in the lane-ownership index). Any code path that shows, hides, silences, or re-lanes a manual lane — a user-created comp/take lane outside the mode system — is a bug. The tool drives only lanes it created.
• Every leaf is managed. The mode system owns all leaf tracks: an untagged leaf is an Arrange member, and when the active mode is not Arrange it is fully parked and snapshot-restored exactly like a tagged leaf out of its mode. show-both is the only way to opt a leaf out of parking. (Parents stay visibility-only, master is never touched — see below.)
• Restore from snapshot, never to a default. No hardcoded "on" restores.
• Verify API names against the SDK header before use.

Open questions to resolve during build

• Reconcile residuals (bulk behavior shipped — ViewModeModel::reconcile(liveGuids) prunes orphaned snapshots on every toggle/load; folder restructure is self-healing because the tree is rebuilt each toggle; membership is intentionally kept so undo-delete preserves the tag). Two sub-items remain deferred:
  • FX-GUID keying for restoreFxOffline: currently restores by slot index; a reshuffled FX chain while parked will restore to the wrong slot. Fix requires FX-GUID keying + snapshot-schema migration — deferred.
  • Dormant membership entries: truly-deleted tracks accumulate stale entries in persisted view_state (harmless and bounded); natural home is a future user-initiated "compact" action, not automatic pruning (which would reintroduce undo-delete tag-loss).
• Interaction with the user having a screenset active (Design View drives the same flags a screenset recall would; last writer wins — confirm no surprising fight).

Two-canvas sub-phase (Phase D2 / Phase E) — additive to D1

Additive sub-phase, settled 2026-07-23. Extends D1's track-level mode projection to item level so each stance owns its own items as well as its own tracks. Nothing in D1 changes; this wraps it. Runtime floor rises to REAPER 7 for this sub-phase (fixed lanes shipped in v7). Product framing in docs/product/design-view.md §Two-canvas direction.

What it adds

On a track present in both stances (a show-both track, or a folder carrying its own media), each mode owns a fixed lane: the active mode's lane shows and plays; the inactive mode's lane is hidden and silent. A Design take and an Arrange take can then live on the same track, same time position, without colliding on the view. Track-only-in-one-mode content is still handled by D1 track-parking, unchanged.

Settled decisions

• Mechanism: fixed item lanes. Map mode → lane; toggle drives per-lane play/show so only the active mode's lane is present. Items keep their real position and real track — nothing is moved in time or deleted. SDK surface (verified present in vendor/reaper-sdk): track-side I_FREEMODE = 2, I_NUMFIXEDLANES, C_LANEPLAYS:N; item-side I_FIXEDLANE, C_LANEPLAYS, B_FIXEDLANE_HIDDEN. I_FREEMODE changes require UpdateTimeline() to take visible effect.
• Membership: auto-tag by active mode at creation. New content — both new tracks and new items — is tagged to whatever mode is active when it is created. Pre-existing content defaults to Arrange. Membership is exclusive: Design-created content never appears in Arrange and vice versa, except via the existing show-both escape hatch. (This is the same tag-to-active-mode rule as D1 track membership, now reaching items on shared tracks.)
• Inactive-mode content is hidden AND silenced. The off-mode lane is set C_LANEPLAYS = 0 — neither shown nor played — consistent with exclusive membership and with D1's "flipping modes is a real change, not cosmetic." Show-both is the deliberate opt-out for a lane that must stay audible across modes.
• Managed vs. manual lanes — indexed and distinct. Fixed lanes are also REAPER's native comping surface: a user may keep their own manual lanes (comp takes, alternate reads) on a track alongside the mode system's lanes. The tool maintains a lane-ownership index — per (track GUID, lane): managed (which mode owns it) vs manual (user-minted, outside the mode system). Mode operations touch only managed lanes; manual lanes are never shown, hidden, silenced, or re-laned by a toggle, and their C_LANEPLAYS stays exactly as the user set it. Items a user adds to a manual lane are not auto-tagged (auto-tag governs normal timeline content, not hand-managed lanes). The ownership index rides in "reasampler" view_state alongside the membership index, GUID-keyed and portable.
• Capture placement is mode-aware. An explicit placement while in Design mode — including capture-and-place — lands the item in the Design lane; the same rule governs manual insertion. The capture load-bearing principle is untouched: capture still writes a file + index entry and never auto-inserts; this governs only where an explicit placement lands.
• REAPER floor: v7. No version-gate branch — below v7 this sub-phase is simply unavailable.

Precision invariants (unaffected — called out explicitly)

The capture precision invariants — null test, bit-identical repeats, non-destructive, exact bounds, relative-paths-only — are entirely unaffected by this sub-phase. No capture path changes; lanes are a placement/view concern downstream of the written file. Lane assignment and C_LANEPLAYS are reversible flags: the item is never relocated in time or deleted, so the sub-phase extends the D1 non-destructive snapshot/restore contract to a new (item-lane) flag family rather than introducing any destructive operation.

New invariant — mode operations touch only managed lanes. A mode toggle drives only lanes the mode system minted (managed lanes, keyed in the lane-ownership index). Manual lanes — user-created comp/take lanes outside the mode system — are never shown, hidden, silenced, or re-laned by a toggle; their C_LANEPLAYS is left exactly as the user set it. This is the fixed-lane analog of never touch B_MUTE/I_SOLO and never touch master: the tool drives only what it created. Enforceable and testable in the pure layer — the "which lanes may this toggle touch" decision is a pure query over the ownership index; only reading REAPER's live lane state is shell.

Module architecture (preserve the pure/shell split)

• Pure (view_mode_model extension). Lane math — which lane maps to which mode, which C_LANEPLAYS value per mode, the item-lane op family alongside the existing track-flag op family — is REAPER-free and unit-tested, mirroring the D1 planner. The lane-ownership index (per (track GUID, lane): managed-which-mode vs manual) and the "which lanes may this toggle touch" query (managed only) are pure and unit-tested — the planner emits lane ops for managed lanes only and never for manual lanes. The auto-tag decision is pure too: given a set of new track/item GUIDs + the active mode — and, for items, whether the item landed in a manual lane (exempt) — produce the membership writes.
• Shell. Two shell responsibilities. (1) Apply the planner's item-lane ops (I_FREEMODE/I_FIXEDLANE/C_LANEPLAYS/B_FIXEDLANE_HIDDEN via the media-item info setters, UpdateTimeline() after I_FREEMODE changes) — for managed lanes only. (2) Detect new content and read live lane state — see below.
• Persistence. The tool persists which lane maps to which mode and the managed/manual ownership index (a small addition to the "reasampler" view section); REAPER stores fixed lanes and lane-plays in the .rpp natively.

New-content detection (implementation design point)

REAPER exposes no clean "item added" / "track added" event callback. Auto-tagging therefore requires the shell to diff project state on the panel's existing timer — the bank_panel already polls and fingerprints the bank; this extends that machinery to the timeline's tracks and items.

• Each poll, compare the live track/item GUID set against the previous poll's set; any GUID new since the last poll is tagged to the then-active mode.
• Correctness the implementation must handle: the first poll after project open must not mass-tag pre-existing content (pre-existing defaults to Arrange, per the membership rule).
• Manual-lane exemption (design point). An item added to a manual lane is not auto-tagged — auto-tag governs normal timeline content, not lanes the user hand-manages. Distinguishing the two may need a heuristic at detection (e.g., an item whose I_FIXEDLANE is marked manual in the ownership index is exempt; content outside any managed lane on a mode-managed track follows the active-mode rule). The precise rule is an open implementation design point; the settled boundary is that manual-lane content is off-limits to auto-tag.
• Lane-identity fragility (design point). I_FIXEDLANE is the lane's identity and is how the ownership index keys to a lane. Whether the index survives lane reorder/renumber/deletion without going stale is an implementation design point (same class as GUID-keyed reorder-safety for tracks) — flag, don't solve here.
• Pure/shell seam: the detection (diffing REAPER's live set each tick, reading live lane ownership) is shell; the tagging decision and the managed/manual lane query (new GUIDs + active mode + manual-lane exemption ⇒ membership + lane writes) are pure and unit-tested.

REAPER API surface (verify all signatures)

• Fixed lanes — track: I_FREEMODE (=2), I_NUMFIXEDLANES, C_LANEPLAYS:N via GetMediaTrackInfo_Value/SetMediaTrackInfo_Value; item: I_FIXEDLANE, C_LANEPLAYS, B_FIXEDLANE_HIDDEN via GetMediaItemInfo_Value / SetMediaItemInfo_Value. Call UpdateTimeline() after I_FREEMODE changes.
• Detection reuses the bank_panel timer + GUID fingerprinting already in place; item GUIDs via the item's GUID (GetSetMediaItemInfo_String "GUID"), track GUIDs via GetTrackGUID as in D1.
• Verify every name/signature against the SDK header before use — the surface is verified present, but confirm argument order, types, and flag values.

---

Multi-bank — additive phase spec

Additive section. This is a standalone phase parallel to — not part of — the M0–M11 capture roadmap and Phase D above. Nothing above changes. Product framing (workflow narrative, pool-privilege reasoning, movement semantics, UI-direction calls) lives in docs/product/multi-bank.md; this section is the authoritative technical spec, matching the house style of the capture and Design View specs. Same standing discipline applies: verify every REAPER API name/signature against vendor/reaper-sdk/sdk/reaper_plugin_functions.h before use.

What it is

The single per-project bank (bank_model / BankIndex) is generalized into a multi-bank system. The existing bank becomes the pool — a default, always-present bank that every capture lands in unless another bank is the active target. On top of the pool the user creates named banks ("Drums", "1-Shots", "Synth Hits") that group samples for a purpose. Samples move freely between any banks, including to and from the pool. Exactly one bank is the active bank — the capture target — the pool by default.

This is the container generalization of the capture pillar's bank. The pool is to banks what Arrange is to modes: structurally one member of an N-collection, but privileged as the default home. The capture pillar's load-bearing rule is untouched — capture still writes a file + an index entry and never inserts into the arrange; the only change is which index the entry lands in.

Settled decisions

• The pool is privileged, not special-cased. Structurally the pool is one BankIndex among many in the container (mirror of "Arrange is just another mode"). Semantically it is privileged: it always exists, is un-deletable, and is un-renamable (fixed id + fixed display name "Pool"). New projects and migrated single-bank projects start with the pool and zero named banks. This keeps the data model uniform (no pool-shaped special type) while the rules layer enforces the three privileges.
• Container in the pure core; a BankIndex per bank. A new pure module bank_book owns an ordered registry of banks, each bank = { stable bank id, display name, ordinal, BankIndex }. BankIndex is untouched — the multi-bank layer wraps it, it does not modify it (additive; no bank-id field on Sample). Bank id is the stable key (GUID-style, minted on bank create); display name and ordinal are mutable (rename / reorder). Display names are unique, enforced in the pure model on create and rename: createBank / renameBank reject a name that duplicates an existing bank's (renaming a bank to its own current name is a no-op success). The comparison is trimmed + case-insensitive (ASCII), so "Drums", "drums", and " Drums " cannot coexist; the pool's reserved name "Pool" is protected by the same check. Uniqueness makes by-name resolution in the action shell unambiguous by construction. The pool is the first, seeded, fixed-id member. bank_book is the mirror of bank_model and view_mode_model: pure, no REAPER types, unit-tested outside the DAW, JSON round-trip.
• Active bank lives in the model, routes through the capture path. bank_book carries the active bank id (defaults to the pool). The capture action family resolves "which bank does this capture land in?" by asking the session for the active bank's BankIndex, then adds exactly as today. No capture backend changes; only the add-target is selected upstream. Activating a bank is a model mutation + a persist write; it never touches the timeline.
• Movement moves the index entry, not the file. Moving a sample from bank A to bank B is an index-only operation: remove the Sample from A's BankIndex, add it to B's. The underlying file stays in the project bank folder — banks are logical groupings over one shared file pool, not separate folders on disk. This keeps movement cheap, non-destructive, and immune to path-rewrite bugs. (Per-bank subfolders on disk are an explicit non-goal — see guardrails.)
• Dedup-by-hash is per-bank. Each BankIndex dedups within itself, unchanged. Moving a sample whose hash already exists in the destination bank collapses onto the existing entry there (the move is a no-op add on the destination side, and the source entry is still removed) — the same collapse semantics BankIndex::add already has, now observed across a move. Cross-bank dedup is not enforced: the same hash may exist in the pool and in a named bank simultaneously (that is the point — copy lets a sample be grouped into "Drums" while still living in the pool).
• Move vs. copy are distinct acts; move is the default. Move removes from source, adds to destination (one logical sample, regrouped) — it is the primary, low-friction gesture, so a sample lives in exactly one bank at a time. Copy adds to destination and leaves the source entry intact (same file, two index entries, two banks) — the deliberate secondary act for the "in two places at once" case. Both are index-only; both share the destination-collapse rule. Under move-as-default the pool is the default home and staging ground, not a permanent superset: moving a sample into a named bank takes it out of the pool. (See product notes for the mental-model reconciliation.)
• Delete drops members; evacuate returns them. Deleting a named bank drops its member index entries (files are not deleted — file lifecycle stays owned by the capture/prune path). A separate evacuate operation moves all of a bank's members back to the pool (index-only, same destination-collapse-by-hash as move), leaving the bank empty. Intended workflow: evacuate then delete to keep the samples, plain delete to drop the grouping and its members. A plain delete of a non-empty bank orphans those members out of every index — their files persist on disk until prune, referenced by no bank — so the UI confirms on non-empty delete and offers evacuate as the alternative. Evacuate cannot be applied to the pool.
• Persistence: a new ext-state key; the pool folds in and the legacy key is retired. The multi-bank state serializes to a new key banks in the existing "reasampler" namespace, alongside view_state and project_guid. The pool's index rides inside the banks blob as bank-zero — persisted identically to any named bank (one blob, one section, one JSON shape). Migration: on load, if a banks key is absent but a legacy bank_index key is present, the legacy index is promoted into the pool inside a freshly-minted banks blob and the book is { pool } with zero named banks — a one-way, lossless promotion. After migration the banks blob is authoritative; the legacy bank_index key is retired (not written or read back going forward). The one-way retirement trades pre-multi-bank backward-read compatibility for the clean single-blob shape — an accepted, forward-only migration consistent with how M4 project state already moves forward.
• Vertical-split UI, pool on top. The bank window splits vertically: pool on top, the named-banks region below (a tab-page strip, one tab per named bank, empty when none exist). Two full-height toggles collapse the split: pool full-height (hide the named-banks region) and banks full-height (hide the pool). The Design View segmented mode switch already in the window header is orthogonal and stays — it governs timeline visibility, not bank grouping; the two coexist in the header/body without interaction.

Precision / invariant implications

• Relative-paths-only survives unchanged. Every BankIndex in the book keeps the relative-path invariant at its add boundary — the book adds no new path handling, because movement is index-only and files never relocate. The invariant is enforced N times (once per bank) by the exact code that enforces it today.
• Non-destructive. Bank create / rename / delete / activate / evacuate and sample move / copy mutate only index + ext-state; no file is written, moved, or deleted, and no timeline item is touched. Deleting a named bank drops its member index entries but does not delete their files; file lifecycle stays owned by the capture/prune path, not the bank container. A file referenced only by the deleted bank becomes an orphan on disk — present but indexed by no bank — until the capture/prune path reclaims it. That orphaned-until-prune window is designed, not accidental; the evacuate verb and the confirm-on-non-empty-delete guardrail exist to keep the user out of it unintentionally.
• Travels-with-the-.rpp preserved. The banks blob rides the same ext-state namespace and the same GUID-primary project-identity / Save-As-relocation machinery as the bank index does today (M4). One shared physical bank folder, one ext-state namespace, now three logical sections (banks + view + identity).
• Determinism / null-test / bit-identical are untouched — they are properties of the capture path and the file, and the multi-bank layer sits above the file entirely.

Module architecture (preserve the pure/shell split)

Pure (no REAPER types, unit-tested — the mirror of bank_model / view_mode_model):

• bank_book — ordered bank registry ({ bank id, display name, ordinal, BankIndex }); pool seeded with fixed id + name; create / rename / reorder / delete named banks (pool-privilege rules enforced here: reject delete/rename of pool; delete drops member index entries; display names unique — create/rename reject a name that duplicates another bank's, trimmed + case-insensitive, "Pool" protected); evacuate a bank (move every member to the pool, index-only, destination-collapse observed; pool cannot be evacuated); active-bank id (get/set, defaults to pool); move and copy a sample between banks (index-only, destination-collapse observed); query a bank's index; JSON round-trip of the whole book (pool-as-bank-zero inside the blob + named banks + per-bank indices + active id + ordinals) and legacy-bank_index→pool migration on parse (one-way; blob authoritative thereafter).

REAPER-facing:

• persist (slice) — serialize/deserialize the book under the banks key in "reasampler" (pool-as-bank-zero inside the blob; no separate bank_index key going forward); migrate a legacy bank_index key into the pool on first load (one-way; blob authoritative thereafter, legacy key retired); reload-on-open and Save-As survival ride the existing M4 machinery. The session exposes the book the way it exposes the bank today; the active bank's BankIndex is what the capture layer adds to.
• bank_panel (extension) — the vertical split: pool grid on top, named-banks tab-page region below; two full-height toggles; the active-bank indicator; the create / rename / delete / activate affordances. The named-banks tab strip is LICE-drawn to match the M5 grid and the Design View segmented switch (not a SWELL-native tab control), with an overflow/scroll affordance so it scales past the ~8–12-tab point. Sample move ships both ways: a "move to bank" menu on the current selection (the bindable front-end for the B3 move action) and drag-between-regions (the direct-manipulation accelerator); copy is the deliberate secondary act, offered on the menu. Drag carries clear drop-target highlighting on the destination region/tab, and a mis-drop is recoverable by design (move is index-only and reversible — the user moves the sample back). Reuses the existing LICE grid render loop per bank region.
• actions (entries) — create bank / rename bank / delete bank (confirm on non-empty delete); evacuate bank → pool; activate bank (direct + cycle); move selected samples → bank; copy selected samples → bank; pool/banks full-height toggles. Registered with the command_id / gaccel / hookcommand pattern; bank-activate + move/copy + evacuate MIDI-bindable to suit the capture-heavy workflow.

REAPER API surface (verify all signatures)

No new REAPER API is invented at spec stage — the multi-bank layer is pure model + persistence + panel UI over machinery M0–M6 already established. Shells will need to verify against the SDK header where they extend existing surfaces:

• Persistence: SetProjExtState / GetProjExtState under "reasampler", new key banks (shared blob machinery from M4 — no new API, new key only).
• Panel UI: the docked-window + LICE-grid surface from M5 (bank_panel), extended to two grid regions + a LICE-drawn tab strip (with overflow/scroll) + toggles. The tab strip, the toggle affordances, and the drag hit-testing are custom-drawn on the M5 LICE surface (not SWELL-native tabs); the "move to bank" menu uses a SWELL popup-menu surface. Verify LICE drawing and any SWELL menu/drag hit-test usage against the M5 reference / SWELL headers, and confirm the drag hit-test does not collide with the M5 grid's multi-select drag. No new REAPER audio API involved.
• Actions: the command_id / gaccel / hookcommand contract from main.cpp (unchanged), new command-id strings under the sampler family prefix.

Non-goals / guardrails

• No per-bank folders on disk. Banks are logical groupings over one shared project bank folder. Do not create a subfolder per bank or move files on bank-move — reject any such path in review (it reintroduces the path-rewrite bug class M4 closed).
• No cross-bank dedup enforcement. The same hash may exist in multiple banks (that is what copy is for). Do not add a global dedup that collapses across banks.
• Pool privileges are inviolable. No action path may delete or rename the pool, leave a project with zero banks, or evacuate the pool (the pool is evacuation's destination, not a source). Enforce in the pure rules layer, not just the UI.
• Delete drops members; files are never deleted by a bank op. Deleting a named bank removes its member index entries only. No bank operation writes, moves, or deletes a file — file lifecycle stays with capture/prune. The UI confirms on non-empty delete and offers evacuate; do not silently orphan members.
• Capture still never inserts into the arrange. The load-bearing principle is unchanged; multi-bank only redirects which index the capture lands in.
• Additive only. Do not alter BankIndex, the M0–M11 capture roadmap, or Phase D semantics. bank_book wraps; it does not modify.
• Verify API names against the SDK header before use.

Open questions to resolve during build

Forks 1–5 are all settled (see product notes → Settled forks and Fork 5 — settled, and the settled-decision prose above). One panel-polish detail remains open.

• Fork 5 — tab rendering + move affordance (B4). Settled (2026-07-23). (5a) The named-banks region is LICE-drawn to match the M5 grid and the Design View segmented switch — not a SWELL-native tab control — with an overflow/scroll affordance in scope from the start so the strip scales past the ~8–12-tab breakdown. (5b) Move ships as both a "move to bank" menu (the precise, MIDI-bindable front-end for the B3 move action) and drag-between-regions (the direct-manipulation accelerator); copy stays the deliberate secondary act via the menu. Drag mis-drop is mitigated by drop-target highlighting and is recoverable by design (move is index-only and reversible). Folded into the bank_panel prose and the API surface below. Verify LICE tab drawing and any SWELL menu/drag hit-test surface against the M5 reference / SWELL headers before use (confirm no collision with the M5 grid's multi-select drag). Analysis in product notes → Fork 5.
• Active-bank indicator placement (B4 polish) — per-region headers vs. a single header readout vs. lit-tab treatment. The "visually unmistakable" requirement is settled (fork 4); only the placement is open. Panel-polish detail.

---

Sample removal — additive spec (Phase B, point B5)

Additive section, part of the Multi-bank pillar. The sample-level companion to move/copy/evacuate/delete-bank: a verb that drops a Sample's index entry from a bank (or the pool). Index-only, non-destructive to the file — it sits on the same side of the index/file line as every other Phase B op. Product framing: docs/product/removal-and-prune.md §Sample-remove. Same verify discipline: verify every REAPER API name/signature against the SDK header before use.

What it is

Move, copy, and evacuate all keep a sample somewhere; there was no verb to drop a sample outright. Sample-remove is that verb: it removes one Sample entry from one BankIndex. It exposes the remove primitive bank_model's BankIndex already has — B5 wires it to an action + a panel affordance, it does not add a model capability.

Settled decisions (spec-level)

• Remove is index-only. It removes the Sample from a BankIndex and mutates only index + ext-state. No file is written, moved, or deleted; no timeline item is touched. Identical non-destructive posture to move/copy/evacuate/delete-bank.
• Remove can orphan a file — the same designed orphaned-until-prune state a non-empty delete-bank produces. When remove drops the last index reference to a file (no other bank holds its hash), that file becomes an orphan on disk, referenced by no bank, reclaimed later by prune (Phase R) — never by remove. This is not a new hazard class; it is the existing "files persist until prune" window, reached by a sample-level verb instead of a bank-level one.
• Collapse-by-hash is unaffected. Remove targets a specific entry in a specific bank. Because cross-bank dedup is deliberately not enforced, removing a sample from one bank leaves any same-hash entry in another bank intact — the same coexistence copy relies on.
• The pool's contents are removable; the pool container is not. Pool privileges (un-deletable, un-renamable, un-evacuable) govern the pool as a container. Individual samples can be removed from the pool — otherwise the pool would be a one-way trap. Remove-from-pool is the pool's own "drop this sample" verb and is allowed.
• Remove scope (fork R-A, SETTLED 2026-07-24 — this-bank). Remove drops the entry from this bank only, leaving copies in other banks untouched — the core and only shipped verb. The action carries a scope: this-bank | all-banks seam, but this-bank is the settled default and the only surfaced affordance; all-banks stays a latent parameter (promotable later behind the seam without a rewrite), never a surfaced verb now. See product notes §Fork R-A.

Precision / invariant implications

• Non-destructive extends to remove verbatim: index + ext-state only, no file touched, no timeline item touched.
• Relative-paths-only is unaffected — remove deletes an entry, it adds no path handling.
• Determinism / bit-identical / null-test (capture) untouched — remove sits above the file, same as all of multi-bank.

Guardrails

• Removes are silent — no confirm dialog. Recoverability is provided by the batched REAPER undo (R-B): one Ctrl-Z restores the index entry, whether or not the sample was a last reference. Files are never deleted by remove (orphaned-until-prune is unchanged). hashReferencedElsewhere is a tested model API retained for Phase R prune; it has no shell caller in the remove path.
• Undo (fork R-B, SETTLED 2026-07-24 — batched REAPER undo points, Phase-B-wide). Bank/index mutations integrate into REAPER's undo system as batched undo points (Undo_BeginBlock / Undo_EndBlock): the related index mutations of one bank operation are batched into a single undo point, so one bank operation is one Ctrl-Z. This is a Phase-B-wide decision — it applies to create/rename/reorder/delete-bank, move, copy, evacuate, and remove, retro- touching B1–B4, not just B5. Must-verify before build: confirm against vendor/reaper-sdk that "reasampler" ext-state mutations participate correctly in Undo_BeginBlock/Undo_EndBlock undo blocks — the whole approach depends on it. Surfaced with remove because remove is the first verb whose only effect is index-entry destruction with no relocation, so it is where the gap first bit; the fix is shared. See product notes §Fork R-B.

Module architecture (preserve the pure/shell split)

• bank_book / BankIndex (pure) — expose remove of a Sample from a bank's index (the existing BankIndex::remove primitive, surfaced through the book); pool contents removable, pool-container privileges unchanged.
• actions (entry) — "remove selected sample(s) from bank" (and, under fork R-A, a scope parameter); registered with the command_id/gaccel/hookcommand contract; MIDI-bindable to suit the capture-heavy workflow.
• bank_panel (affordance) — remove on the current selection (menu entry / key), reusing the M5 selection model exactly as move/copy do; confirm-on-last-reference at this layer.

REAPER API surface

No new REAPER API. Pure model + a new action command-id string under the sampler family prefix + a panel affordance on the existing M5 LICE surface. Verify the command-id/gaccel/hookcommand usage against main.cpp (unchanged contract).

Settled forks (Daniel, 2026-07-24)

• Fork R-A — remove scope. Settled: this-bank (this-bank-primary, all-banks a latent seam-only parameter). Folded into Settled decisions above.
• Fork R-B — undo model for index mutations. Settled: batched REAPER undo points (Undo_BeginBlock/Undo_EndBlock), Phase-B-wide (retro-touches B1–B4), with the ext-state-participation SDK check as a must-verify-before-build. Folded into Guardrails above and the Phase B / B1 plan points.

---

Prune — file-lifecycle spec (Phase R — Reclaim)

New pillar, its own lettered phase. Prune is the file-lifecycle path the capture and multi-bank specs forward-reference throughout ("files persist on disk until prune", "the capture/prune path reclaims it") but that had no phase, module, or point until now. It is the only operation in ReaSampler that deletes bytes off disk. Namespaced R (Reclaim) alongside M (capture), D (Design View), B (Banks) — it is a distinct pillar, not a Multi-bank sub-step, because it serves every orphan-producing path (delete-bank, sample-remove, re-capture) and carries a new risk class (file deletion) with its own invariants. Product framing and the phase-placement justification: docs/product/removal-and-prune.md §Prune. Same discipline: verify every REAPER/SWELL/filesystem API name/signature against the SDK/SWELL headers before use.

What it is

Over a project's life, delete-bank and sample-remove (and, potentially, M10 re-capture superseding an old file) leave .wav files on disk that no bank index references — the "orphaned-until-prune" state the specs design in on purpose. Prune is the reclaim pass: reconcile the physical bank folder against the union of every bank's index, and reclaim the files nothing references. It makes good on the promise the rest of the spec keeps making.

The load-bearing rule

Remove creates orphans; prune reclaims them. Sample-remove and delete-bank drop index entries and may leave a file referenced by nothing. Prune is the single path that turns such an orphan back into free disk space. No other operation deletes a file; prune deletes only files that no index references. A bank op that deletes a file is still a bug — prune is not a bank op, it is the file-lifecycle op.

This asymmetry is deliberate and must be stated loudly: every other invariant says "no operation deletes a file." Prune is the sole, explicit exception, and its entire job is deletion — so it must be the only file-deleting authority in the system, with the strongest guardrails.

Mirror of reconcile — the pure pattern one level down

Prune reuses the shape Design View already shipped. view_mode_model's ViewModeModel::reconcile(liveGuids) reconciles membership entries against live tracks and returns the residuals to drop. Prune reconciles files on disk against referenced files (the union of every bank's index) and returns the orphan set to delete. Same pure pattern, one level down (files instead of GUIDs).

The decision is pure and unit-tested: given the set of files present in the bank folder and the set of files referenced by the book, compute the orphan set. Only the two ends touch the shell — enumerating the bank folder and deleting the orphans are filesystem I/O. Keep the "which files are orphans" core REAPER-free and hard-tested (this is the safety-critical part); keep the I/O thin. Same pure/shell split as bank_model / view_mode_model / bank_book.

Settled decisions (spec-level)

• Referenced-set is the union across ALL banks, pool included. A file is an orphan iff no bank in the book references it. Because copy lets one file be referenced by several banks, prune must union references across the whole book before deciding. This is the safety-critical computation — the prune null test is prune never deletes a file that any index references.
• Project-relative resolution, current folder. Prune enumerates and deletes within the project bank folder using the same M4 project-relative path resolution the index uses, against the resolved current folder — never a stale absolute path — so a Save-As relocation cannot cause it to mis-identify or mis-target orphans.
• Dry-run first, always. Prune reports before it deletes: the orphan count, reclaimed size, and (for a small set) the files. The dry-run — compute-and-report, the pure core with no deletion — is the primary surface; actual deletion is the confirmed second step. A prune that silently sweeps is unacceptable for an irreversible file-delete.
• Scope is the bank system's own leavings, not the folder at large. Prune reclaims files that were bank files and are now unreferenced — never a file a user hand-dropped into the folder. Prune is a reclaimer of ReaSampler's own orphans, not a general folder cleaner.
• Orphan attribution is an owned-file manifest (fork R-D, SETTLED 2026-07-24). The book tracks the set of files it has created (an owned-file manifest); prune reclaims (owned ∩ on-disk) − referenced. This is the honest encoding of "reclaim only our own leavings" and rejects folder-sweep (which would delete hand-dropped files). The seam lands early: because the manifest is cheap to maintain from capture onward but a backfill cliff to reconstruct later, capture writes each file it creates into the owned-file manifest starting in Phase B, even though prune consumes it only in Phase R. R1/R2 consume the manifest; they do not build it. The manifest is persisted in the "reasampler" ext-state; the exact persistence shape (a sibling key vs. folded into the banks blob) is a small build-time residual, but the manifest-now decision is firm.

Precision / invariant implications

• The single intentional exception to "no operation deletes files." Stated above; called out again here so the invariant table is honest: prune is destructive-to-files by design and by exclusive authority.
• Relative-paths-only / Save-As machinery reused — prune resolves paths the same way the index does (M4), so it inherits relative-path correctness and Save-As survival; it introduces no new path handling.
• Determinism / bit-identical / null-test (capture) untouched — prune sits below the capture path entirely.
• Prune null test (new invariant): a prune of a folder whose every file is referenced by some bank deletes nothing; a prune deletes exactly the present − referenced orphan set and nothing else. Ship as a tested property of the pure core.

Guardrails — the genuinely destructive act

• Dry-run + confirm-with-manifest (above): the user approves a specific deletion (count + size + files), never an abstract "clean up."
• Never a referenced file; never a non-bank file. The union-across-all-banks rule protects referenced files; the ownership-attribution rule (fork R-D) protects hand-dropped files.
• Safest platform deletion available (fork R-C, SETTLED 2026-07-24 — trash- preferred, unlink fallback). Route deletions to the platform recycle bin / trash wherever a portable move-to-trash is available (recoverable outside the app); fall back to unlink — behind the dry-run + confirm guardrail — only where the platform affords no portable trash. "Delete where possible" means recoverable-trash- preferred, never plain unlink-by-default. The move-to-trash surface is an explicit per-platform to-verify (see REAPER/platform API surface). (R3 verified: Windows SHFileOperationW + FOF_ALLOWUNDO confirmed against SDK 10.0.26100; macOS/Linux unlink fallback — no portable SWELL trash surface.)
• Manual, explicit trigger (fork R-E, SETTLED 2026-07-24 — manual action + panel button). Prune runs via a bindable manual action (dry-run-first, confirm-to- delete) and a bank_panel button that fires that same action — never a silent background sweep. The earlier optional "…and prune now at the delete-bank confirmation" convenience was not selected and is out of scope; a periodic background sweep remains rejected (silent irreversible file-deletion violates the guardrails).

Module architecture (preserve the pure/shell split)

Pure (no REAPER types, unit-tested — the mirror of reconcile):

• Prune-reconcile core — given { files present in the bank folder }, { files referenced by the book }, and { files the book owns } (the owned-file manifest, R-D), compute the orphan set (owned ∩ present) − referenced. REAPER-free, filesystem-free, unit-tested hard (the prune null test lives here). The referenced-set is unioned across all banks by asking the bank_book.

REAPER-facing / filesystem-facing (thin):

• persist / session — supplies the referenced-set (union across the book) and the owned-file manifest (R-D, written from capture onward in Phase B); resolves the current project bank folder via the M4 project-relative machinery.
• A prune shell — enumerates the bank folder (filesystem I/O), feeds the pure core, presents the dry-run manifest, and on confirmation deletes the orphan set (via OS trash where portably available — fork R-C — else unlink). Filesystem I/O only; the decision stays in the pure core.
• actions (entry) — "Prune bank folder" (dry-run-first, confirm-to-delete), registered with the command_id/gaccel/hookcommand contract; plus a bank_panel button (R-E) that fires the same action.

REAPER / platform API surface (verify all signatures)

No new REAPER audio API. New surfaces to verify before use:

• Filesystem enumeration + delete — directory listing and file removal for the project bank folder. Verify the portable approach against SWELL / the existing file-handling in persist / capture (which already resolve and write files); prefer reusing whatever path/file machinery M4 established.
• Move-to-trash (fork R-C, settled trash-preferred) — verify a portable move-to-trash exists (SWELL, or per-platform: Win IFileOperation/ SHFileOperation, macOS NSFileManager trashItemAtURL:, Linux XDG trash spec). This is a must-verify per platform before use, not an assumed capability; where it is unavailable, fall back to unlink behind the dry-run/confirm guardrail. (R3 verified: Windows routes to Recycle Bin via SHFileOperationW + FOF_ALLOWUNDO, verified against SDK 10.0.26100. macOS / Linux: no portable SWELL trash surface found — fall back to unlink behind the dry-run/confirm guardrail, as specified.)
• Owned-file manifest persistence (fork R-D, settled) — a new tracked set in the "reasampler" ext-state (a sibling key or folded into the banks blob — build-time residual); shared M4 blob machinery, new data only. Written from capture onward in Phase B (the seam lands early), consumed by prune in Phase R.
• Actions — the command_id/gaccel/hookcommand contract from main.cpp (unchanged), a new command-id string under the sampler family prefix.

Non-goals / guardrails

• Prune is the ONLY file-deletion authority. No bank op, no capture op, no Design View op deletes a file. If any path other than prune deletes a bank file, reject it in review.
• No general folder cleaning. Prune reclaims the bank system's own unreferenced leavings, not arbitrary files a user placed in the folder (fork R-D governs the attribution).
• No silent deletion. Dry-run + explicit confirm always; no background sweep.
• No file deleted while any index references it. The referenced-set union across all banks is the safety-critical invariant — enforce and test it in the pure core, not just the UI.
• Additive only. Prune reads the book and the folder; it does not modify BankIndex, bank_book, the capture roadmap, or Design View semantics.

Settled forks (Daniel, 2026-07-24)

• Fork R-C — deletion mechanism. Settled: trash-preferred, unlink fallback. Route to the OS trash where a portable move-to-trash is available (recoverable), else unlink behind the dry-run/confirm guardrail. Per-platform trash surface is a must-verify. Folded into Settled decisions + Guardrails + API surface above. Product notes §Fork R-C.
• Fork R-D — orphan attribution. Settled: owned-file manifest, (owned ∩ present) − referenced; folder-sweep rejected as unsafe. The seam lands early — capture writes each created file to the manifest starting in Phase B, prune consumes it in Phase R. Persistence shape (sibling key vs. banks blob) is a build-time residual. Folded into Settled decisions + Module architecture + API surface above, and added as an up-front Phase B / capture plan point. Product notes §Fork R-D.
• Fork R-E — trigger. Settled: manual action + bank_panel button, dry-run- first, confirm-to-delete; no background sweep. The delete-time "…and prune now" convenience was not selected (out of scope). Folded into Guardrails + Module architecture above and the R3 plan points. Product notes §Fork R-E.

Build-time residual (not a fork): the owned-file manifest's exact persistence shape (sibling "reasampler" ext-state key vs. folded into the banks blob).

---

MIDI-playback instrument — additive phase spec (Phase S — Sampler)

New pillar, its own lettered phase, and — uniquely — its own build artifact. Every prior phase (M / D / B / R / V) ships inside the one reaper_reasampler extension binary. Phase S does not: a REAPER extension cannot be a MIDI-triggered instrument (it is not a node in any track's signal chain), so the instrument is a second, separate binary — a native VST3 plugin the user instantiates on an instrument track — that reads ReaSampler's banks and plays them MIDI-triggered. Namespaced S (Sampler) rather than "D" (which would collide with Design View). The M/D/B/R/V extension pillars are untouched. Product framing, the plugin-format reasoning, the bare-VST3-vs-JUCE assessment, and the settled decision record: docs/product/midi-playback.md. Same standing discipline: verify every Steinberg VST3 SDK and REAPER/SWELL API name/signature against the vendored headers before use.

What it is

A native VST3 sampler instrument — a separate product/artifact from the extension — that maps ReaSampler's captured bank samples across a MIDI keyboard and plays them back with a real voice engine (polyphony, velocity, envelopes). The extension stays the sole owner of capture + organization; the instrument is the playback surface. The two are tightly integrated but distinct acts: the extension captures and organizes; the instrument plays. Neither crosses into the other's role — the instrument never captures, the extension never becomes an instrument.

Why a VST3 and not the extension (load-bearing, settled — see D1/D5/D6 below). An instrument track's "read live MIDI, emit audio per-voice, in REAPER's routing/record/render path" contract belongs to VST/VST3/CLAP/JSFX plugins, hosted through an entirely different mechanism than the extension API. The extension SDK's audio-adjacent surfaces (Audio_RegHardwareHook, kbd_OnMidiEvent, PlayPreview, pcmsrc subclassing) are each the wrong tool for a live-MIDI instrument — the full reasoning is in docs/product/midi-playback.md §1. The instrument is therefore a standard VST3 plugin; this is not an engineering-around-able limitation, it is what the plugin format is.

The three locked decisions this spec assumes

Settled by Daniel (2026-07-26); everything below assumes them. Reasoning preserved in docs/product/midi-playback.md §4.

• D1 — native VST3. Not JSFX. Full sampler sophistication, clean integration, and access to the REAPER VST-host bridge. JSFX retired (cross-platform-for-free is worthless under D5, and JSFX gets no bridge).
• D5 — Windows-only, VST3-only, REAPER-only. No cross-platform DSP/build/signing matrix, no multi-format wrapper, no standalone-in-other-hosts concern. REAPER-coupling via the bridge is intended. This is the single biggest simplifier — it deletes most of what makes VST3 painful.
• D6 — two products, tightly integrated. A separate artifact, but not a divorced file-only companion: via the VST-host bridge it reads the live "reasampler" project ext-state and is project-aware.

The VST-host bridge (the integration mechanism, stated once)

A VST3 hosted inside REAPER can call back into REAPER's own API by resolving function pointers by name over the host callback (hostcb(&effect, 0xdeadbeef, 0xdeadf00d, 0, "FunctionName", 0.0) — the same string-keyed API table the extension uses; verified in video_processor.h and the reaper_plugin_functions.h GetProjExtState/SetProjExtState/EnumProjExtState entries). The plugin can also fetch its host context — the track/take/project it was instantiated in (opcode 0xdeadf00e). Consequence: the instrument reads the same live "reasampler" ext-state that persist writes, follows the active project, and needs no "point me at the bank folder" wiring — it asks REAPER which project it is in. This capability exists because the plugin is hosted in REAPER; it is the technical affordance D6 leaned on. Must-verify before build: confirm the bridge opcodes and the by-name resolution against the vendored vendor/reaper-sdk/sdk/ headers (reaper_plugin.h, video_processor.h, reaper_plugin_functions.h) — the framing doc's opcode citations are verified from those headers but the exact call marshalling should be confirmed at the spike.

The two seams (audio via files, mapping via live state)

• File seam (audio, permanent). The sample audio is the on-disk 32-bit-float WAVs — project-relative, travelling with the .rpp via the M4 machinery. The instrument resolves those paths the same way persist does (a shared convention, not a re-implementation). There is no live PCM stream across the bridge, by design.
• Live-state seam (the mapping, via the bridge). For everything that is not raw audio — the bank index, the mapping, which project is active — the instrument reads the live "reasampler" ext-state via the bridge. It sees what persist last wrote and follows the active project.

The seam fields — what becomes a bank intrinsic (D-B, settled 2026-07-26)

The split model (option iii) is the settled answer. It mirrors the capture/placement separation:

• Bank intrinsics (facts about the captured file) live on Sample. Root note (the MIDI note the sample was recorded at, so it can be repitched across the keyboard — distinct from the existing optional musical key field) and loop points (sustain-loop start/end for held notes; sample-accurate, zero-crossing-aware) are facts about the file, analogous to sample rate, length, and peaks. They are added to Sample as an additive field extension — the same shape as how provenance was added in M1: new optional fields with JSON round-trip, populated at/after capture, defaulting cleanly for pre-existing samples. This keeps the bank a clean, tool-agnostic library (WAVs + facts, readable by anything).
• The performance map (a creative arrangement) lives in the instrument. Key zones (low/high note per sample), velocity layers, round-robin groups, amplitude envelopes (ADSR), and per-sample tuning/gain trim are a performance choice, not a fact about a file — they belong to the instrument, not the bank. Under the live-state seam the instrument may still read performance-map data out of shared "reasampler" state, so "who owns which field" is a data-ownership decision, not a transport one.

Why the intrinsic fields are added early (D-B, the backfill-cliff reasoning). The Sample field addition is scheduled as an early Phase S point even though the instrument that consumes them lands later. Rationale (the design-the-seam-even-if-you- defer-the-feature instinct, same as Fork R-D's owned-file manifest): if the fields are added only when the instrument needs them, every sample captured before then lacks a root note / loop points and must be backfilled by hand. Adding the fields now — so capture starts populating them (or at least defaulting them cleanly) — costs almost nothing and closes the cliff. The field addition touches the extension codebase (bank_model + capture + persist), is independently shippable, and lands before the instrument build leans on it.

Scope tiers (D-C, settled 2026-07-26 — Tier 0–1 committed, Tier 2 held, Tier 3 optional-forever)

Tiers are minimal → sophisticated; Tier 0 delivers the core promise and each tier above is optional depth, not a prerequisite for the one below.

• Tier 0 — "the bank plays" (committed). One sample mapped chromatically across the keyboard from its root note; basic polyphony; a simple amp envelope; velocity → volume. The honest MVP: point a bank sample at a MIDI track and play it repitched. On the native path this is the SingleComponentEffect skeleton plus a single-voice core, editor deferrable behind a parameters-only default view.
• Tier 1 — "a keymap" (committed). Multiple samples zoned across the keyboard (key ranges), each with its own root note — a captured kit or a multisampled instrument plays correctly. This is where the root-note + key-range seam fields earn their place. One sample per key-region.
• Tier 2 — "expressive" (HELD — noted, not specified). Velocity layers, round-robin (the anti-machine-gun feature), full ADSR, per-sample tuning/gain trim, sustain loops. Where it becomes a tool people reach for. Explicitly a follow-on — its points are not drawn up in this spec; it is recorded as the next depth increment once Tier 0–1 proves the instrument belongs in ReaSampler's world.
• Tier 3 — "instrument polish" (optional-forever). Filters, filter/pitch envelopes, LFOs, per-voice pan, choke groups, a modest FX slot. A direction to leave room for, never a commitment. Do not let a Tier-3 feature list inflate the build-shape decisions.

The build shape (D-A, settled 2026-07-26 — bare Steinberg VST3 SDK + LICE editor)

Settled: bare Steinberg VST3 SDK, no JUCE, with the editor drawn in the same LICE/SWELL stack bank_panel already uses. Reasoning (full assessment in docs/product/midi-playback.md §1a and §4 D-A):

• The audio-processing scaffolding is bounded. Using SingleComponentEffect (the SDK's combined processor+controller base — sanctioned for a non-distributable, REAPER-only plugin under D5/D6) plus the SDK's factory macros, a silent-but-loading VST3 instrument skeleton is order-of-magnitude a few-hundred lines of adapt-from-example ceremony, written once. The AGain / Note Expression Synth SDK examples are the copy-source. Not a tar pit.
• D5 deletes JUCE's biggest justification. JUCE exists largely for multi-format / cross-platform, both of which D5 removed. Its one genuine remaining pull is the editor UI — and ReaSampler is the atypical case where even that is weak, because it already has a working, docked, custom-drawn LICE UI (bank_panel) and a house style. Drawing the editor in a VST3 IPlugView that hosts a LICE surface reuses that muscle, keeps the look house-consistent, and avoids JUCE's AGPL-or-pay license posture (the Steinberg SDK is permissive, no revenue gate).
• The one real edge — the IPlugView↔LICE bridge (window lifecycle, sizing, event routing from the host into the draw/hit-test loop) — is the same class of work ReaSampler already did to dock bank_panel, not a new competence, but it is less trodden than dropping in a JUCE editor. It is therefore the phase's opening spike (below), which also converts §1a's experienced-estimates (Windows module-export symbol names, factory-macro spellings, exact bridge marshalling) into verified fact before the engine build leans on them. VSTGUI (the SDK's bundled toolkit) is the noted fallback rung only if the LICE bridge proves gnarlier than the panel work suggests; JUCE is the last resort behind that.

The pure core (D3 — the load-bearing split, transplanted)

The sampler's voice engine, envelope math, key/velocity mapping, repitch/interpolation, and keymap resolution are a pure, REAPER-free, DAW-free, unit-tested module — the mirror of bank_model / peaks / view_mode_model / bank_book, tested in the CTest harness outside any host. This is the heart of the phase; test it hard. The VST3 wrapper — the SingleComponentEffect subclass, bus setup, the process call marshalling MIDI→core and core→audio-buffer, the IPlugView LICE editor, and the bridge calls that read "reasampler" ext-state — is the thin shell, the only part that touches VST3 or REAPER at all. Critically, this split is invariant under the build-shape choice: whether the shell is bare-SDK or (hypothetically) JUCE, the pure core is identical, REAPER-free, and tested the same way. The format choice is a shell choice; the core is invariant.

Module architecture (preserve the pure/shell split — in the new artifact)

Pure (no REAPER types, no VST3 types, unit-tested — the mirror of bank_model):

• Sampler core — voice allocation/polyphony, amplitude envelope (ADSR), key→sample and velocity→sample mapping (the keymap), repitch/interpolation from root note, and keymap resolution. REAPER-free and VST3-free, unit-tested in CTest against known signals (mirror of how peaks asserts an envelope). This is D3's pure core and the heart of the phase.

Shell (VST3-facing / REAPER-facing, thin):

• VST3 wrapper — SingleComponentEffect subclass: initialize (declare an event input bus + an audio output bus, no audio input), setupProcessing, setActive, setState/getState, and the hot-path process that reads MIDI off the event bus, drives the pure core, and writes the core's per-voice audio to the output bus. Plus the module factory (GetPluginFactory + Windows InitDll/ExitDll — verify exact export names at the spike).
• IPlugView LICE editor — hosts a LICE-drawn surface in the VST3 view seat (window creation/sizing, host→draw/hit-test event routing). Reuses the bank_panel LICE/SWELL competence and house style.
• Bridge/state reader — resolves GetProjExtState/EnumProjExtState by name over the host callback, fetches the host project context, reads the live "reasampler" ext-state (bank index + intrinsic fields + performance map), and resolves WAV audio paths the same project-relative way persist does.

Precision / invariant implications

• The bank is one source; the instrument is another view of it (never a fork). The instrument is a pure consumer of the bank — it does not copy samples, does not own a private sample store, and does not mutate the bank. The bank stays the single authoritative artifact (the one-source-multiple-views instinct). Any instrument path that writes back into the bank or keeps its own sample copies is a bug.
• Capture/placement/playback stay distinct acts. The instrument reads and plays; it never captures and never inserts into the arrange. The capture load-bearing principle is untouched — Phase S adds a third distinct act (playback) without weakening the capture↔placement separation.
• Sample field addition is additive and lossless. Root note + loop points are new optional fields with JSON round-trip, defaulting cleanly for samples captured before the addition — the same additive, backward-compatible shape as provenance (M1). No existing Sample field changes; no BankIndex behavior changes.
• Relative-paths-only survives. The instrument resolves audio via the M4 project-relative machinery; it introduces no absolute paths.

Embedded TCP/MCP UI (D-D, settled 2026-07-26 — SCHEDULED as a later Phase S point)

A REAPER-hosted VST3 can draw its own UI inline in the track/mixer control panel via reaper_plugin_fx_embed.h (the plugin implements IReaperUIEmbedInterface; the same Cockos surface REAPER's own embedded FX use). A ReaSampler instrument can render a compact keymap/level strip inline in the TCP/MCP, not only in its own window. Because this uses the same LICE-class drawing as the D-A editor path, it composes naturally with the bare-SDK-plus-LICE build — the groundwork is the groundwork.

Settled: scheduled, not deferred. This is a real, in-phase later point — it lands after the main IPlugView editor exists (it composes with that LICE path), not a someday-note. It is polish, not a Tier-0 need, so it sequences last in the phase; but it is on the roadmap. Must-verify before build: the IReaperUIEmbedInterface contract and embed message/lifecycle against vendor/reaper-sdk/sdk/reaper_plugin_fx_embed.h.

Channel mode — mono | stereo (D-E, decided 2026-07-26; PLAN.md S7)

Decided direction: the instrument gets a per-instance channel-mode toggle — 1 (mono) or 2 (stereo) — that negotiates the REAPER audio bus automatically. Captures are often stereo; the current mono downmix is a Tier-0 simplification, not a permanent shape.

• Mono mode keeps today's path. The decode-side downmix stands: a stereo source in mono mode downmixes (the existing policy), a mono source plays as-is. No engine change for mono.
• Stereo mode is an S3-core extension, not a shell hack (honest). The S3 core is mono-per-sample by design today — SampleData::frames is one mono stream, Voice::renderFrame returns a single value, VoiceEngine::render writes one channel. Stereo mode grows the core a channel dimension: 2-channel decoded PCM, per-voice stereo render (per-channel fractional read + linear interpolation + loop), and a per-channel mix in the engine. Mono stays the degenerate (single-channel) case, so existing mono behavior is unchanged. This is why S7 sequences first after the editor/embed work: it touches the engine Daniel smoke-tests.
• Where the toggle lives. Per-instance component state (setState/getState), alongside the selected sample — a performance choice the instrument owns, never written to the bank (D-B: not a file fact). Default preserves current behavior (mono).
• Cross-mode policy (settled). Mono source + stereo mode → dual-mono (same signal both channels, centered). Stereo source + mono mode → downmix (the existing decode-side policy). The bank's per-sample channel-count intrinsic (already on Sample) tells the shell how many channels to decode into SampleData.
• Bus negotiation (the "works with the REAPER bus automatically" requirement). The VST3 implements setBusArrangements so the output bus reports mono or stereo per the instance's channel mode, and REAPER's routing follows without manual channel wiring. Must-verify before build: the setBusArrangements / getBusArrangement contract and REAPER's mono/stereo instrument-bus expectations against the vendored Steinberg SDK + reaper_vst3_interfaces.h.

Sampling modes — Trigger vs Gate + pitch envelope (S15/S16; core + editor)

Daniel's directive (2026-07-26, verbatim): "Sampling mode: Trigger vs Gate. Gate has an AHDSR envelope. Trigger has fade in, % length, and fade out. Both modes have modifiable start point, Gate has modifiable loop points too. In addition to amp env, there will be a pitch envelope/curve (AD?) which is off by default." The feature set is settled; two forks (S15-F1 choke, S15-F2 param granularity) are flagged with leans below.

Play mode — Gate vs Trigger (S15)

Each played sample carries a play mode — a per-sample/per-zone performance choice (D-B, instrument-owned, never a bank fact). Two modes, precisely:

• Gate — classic held note (grows the current path). Note-on enters the amp envelope; note-off enters release; a sustain loop applies for held notes (S11's draggable loop markers are Gate-mode UI). The current core envelope is ADSR; Gate adds a Hold stage → AHDSR: 0→1 over attack, hold at 1 over holdFrames, 1→sustain over decay, hold sustain until note-off, level→0 over release. holdFrames == 0 is exactly today's ADSR — a back-compat degenerate, no behavior change for existing Gate play. Segment math is the existing linear-ramp idiom (AdsrEnvelope::tick) with one new stage inserted between Attack and Decay.
• Trigger — one-shot drum-pad. Note-on fires playback of a defined % of sample length with a fade-in and fade-out ramp; note-off is ignored (the voice plays through); no sustain loop. Envelope math (distinct from AHDSR): play the frame span [startFrame, playEnd) where playEnd = startFrame + round(lengthFraction·(frames − startFrame)), lengthFraction ∈ (0,1]; amplitude ramps 0→1 over fadeInFrames (fade-in) at the head and 1→0 over fadeOutFrames anchored to playEnd (fade-out), unity between; fades clamp so fadeInFrames + fadeOutFrames ≤ play length. The voice frees when readPos_ ≥ playEnd (mirror of the current run-off-end idle). Fade curve default: equal-power (constant-power sin/cos — click-free on one-shots); linear is a build-time residual. Note-off in Trigger is a no-op (choke is held — fork S15-F1).

Both modes: modifiable start point. Playback begins at startFrame (a frame offset into the sample, clamped 0 ≤ startFrame < frames), not always frame 0. This is the voice's initial readPos_; the existing per-frame readPos_ += ratio_ read and linear-interp/loop machinery are otherwise unchanged. Gate additionally has modifiable loop points (already the S2 loop intrinsic + S11 override); Trigger has none (it is a one-shot).

Voice-stealing interaction (unchanged). The S3 stealing policy (oldest-in-release, else oldest-overall) is mode-agnostic — a Trigger one-shot is a normal active voice until it runs off playEnd; it can be stolen like any voice. No new stealing rule.

Confirmed from the core (sampler_core.cpp): the read loop advances readPos_ by an arbitrary ratio_ per frame with 2-point linear interpolation, and the amp is a per-frame env_.tick() multiply — so both the AHDSR hold stage and the Trigger fade/%-length envelope are per-frame amplitude functions over the existing read machinery, and the start point is just a non-zero initial readPos_. No resampler or voice-lifecycle rewrite is needed.

Parameter ownership (D-B). The play mode + its params (Gate: AHDSR; Trigger: %-length + fade-in + fade-out; both: start point) attach to the capture selection / zone and live in the instrument's performance map (component state, version-bumped, back-compat: a truncated/older blob defaults to Gate, hold=0, start=0, no fades = exactly today). Start point joins rootOverride / loop-override as another per-PerformanceZone optional override; a per-zone PlayMode + param struct is added additively. Fork S15-F2 (flagged): per-capture-selection and per-zone, or per-zone only with the single-capture case as a one-zone map? Lean: per-zone only — the single capture is already a one-zone map (S10-Z's back-compat lift), so one storage site serves both; flagged because it touches S10's single-capture setup surface shape.

Editor (mode-aware, on the S11 waveform surface). Gate shows draggable start + loop markers; Trigger shows start + %-length end + fade-in/out handles — same waveform, same pure frame↔pixel + marker-grab geometry module (S11), mode switches which markers draw. A mode toggle per capture/zone sits in the S10 guided setup / S10-Z Zones panel. Every edit commits off-thread via commitMapAndReload; the instrument stays a read-only bank consumer (mode/params are performance map, never written to Sample or the bank).

Pitch envelope — AD, off by default (S16)

A per-voice pitch modulation curve on top of a zone's base repitch — a short AD (attack-decay) envelope that biases the voice's read-increment over time. Off by default (so existing playback is bit-identical). The classic use is a percussive pitch drop.

• Shape (lean, build-time residual): two-segment AD — at note-on the pitch offset rises to peakSemitones over attackFrames, then falls to 0 (base pitch) over decayFrames. A zero attack gives the pure "start high, drop to base" percussive drop; a positive peak that settles to 0 is the classic sampler pitch envelope. Documented; the alternative (start-offset → monotone glide to base) is the degenerate attack=0 case of this.
• Range: semitones (±). peakSemitones is signed (positive = start/peak above base, negative = below); default depth range noted at build.
• RT implication — confirmed clean. The core's resampler is already an arbitrary per-frame ratio linear interpolation (readPos_ += ratio_, verified in sampler_core.cpp). The pitch envelope is therefore a per-frame multiply of ratio_ by 2^(pitchEnvSemitones(frame)/12) — the effective read increment varies frame-by-frame at no structural cost. No new resampler, no WDL dependency for the modulation path. The envelope is the same per-frame tick() idiom as the amp envelope — RT-safe, no allocation in process, per-voice (polyphonic notes each run their own).
• Ownership + editor. Per-zone instrument performance-map state (D-B), additive/version- bumped (absent → disabled). Editor exposure folds into the S12 ADSR-editor tier: attack + decay + a ±semitone depth control, default-off (discoverable but inert until enabled).

WDL pitch/resample surface — verified finding (feeds S15/S16, not a committed point)

The full vendored WDL pitch/resample surface was swept — vendor/WDL/WDL/resample.h and vendor/WDL/WDL/simple_pitchshift.h are the only two pitch/resample headers; there is no elastique / formant-preserving / time-stretch anywhere in the tree. Honest findings:

• WDL_Resampler (resample.h) — sinc/linear resampler, RT-suitable. SetMode(interp, filtercnt, sinc, sinc_size≤64, sinc_interpsize); streaming ResamplePrepare/ResampleOut with Prealloc. Its sinc mode beats the core's 2-point linear interp for base-repitch quality (less aliasing on large transpositions) at a real CPU cost (up-to-64-tap convolution per output sample vs. one lerp). Fit: an optional quality upgrade for the base repitch path (a per-voice linear/sinc toggle) — not required for S15/S16 and not committed; held as a Tier-2/3 quality option (per-voice WDL_Resampler instances are heavier, and S16's modulation is cleaner hand-rolled).
• WDL_SimplePitchShifter (simple_pitchshift.h) — time-domain OLA pitch shifter, wrong tool. It shifts pitch preserving duration — the opposite of a sampler's repitch-by-resampling (which changes pitch and duration together). Its set_formant_shift is an explicit empty stub (no formant preservation). Not a fit for the sampler repitch/envelope path; noted for completeness.
• Formant-preserving / time-stretch (elastique-class): NOT available. REAPER's elastique is licensed (zplane), not in the open WDL/reaper-sdk vendored tree (grep found only unrelated libpng/giflib string matches). Formant-correct repitch is unavailable to the instrument without a new third-party dependency — out of scope (D5-adjacent). Stated, not worked around.
• Recommendation: S16's pitch-envelope ratio-modulation stays hand-rolled (per-frame ratio_ multiply over the existing linear-interp read — simplest, RT-safe, already supported). WDL_Resampler (sinc) is the only WDL piece worth adopting, and only as an optional base-repitch quality upgrade — held, not scheduled.

Sequencing (S15/S16 against S7 stereo, S10 editor)

S15 and S16 are S3-core extensions — they touch the engine Daniel smoke-tests, like S7. They are channel-count-agnostic by construction: the play-mode envelope is a per-frame amplitude function and the pitch envelope is a per-frame read-rate scalar — both independent of how many channels SampleData carries. So they compose cleanly with S7's channel dimension rather than conflicting: S7 adds a channel axis to the read/mix; S15/S16 add an amplitude-shape axis and a read-rate axis; the three are orthogonal. Recommended order: S15 before S16 (S16's pitch envelope reuses S15's per-voice param-plumbing + component-state version bumps; landing S15's PlayMode/param struct first gives S16 a home to hang the pitch-env params on). S15/S16 relative to S7: no hard dependency — spec them so the envelope/mode code never assumes a channel count (it operates per-frame, pre-mix), and S7 can land before, after, or interleaved. Relative to S10 (editor): S15's mode toggle + Trigger handles and S16's AD control surface through the S10/S11 waveform + guided-setup work, so the core halves of S15/S16 can land independently of the editor, with the editor surfacing following S10/S11 (the same way S12's ADSR editor follows the S3 ADSR math). Land the core engine work (mode split, start point, %-length/fades, pitch-env modulation) as soon as it is ready — it is testable in CTest without the editor — and wire the UI as the S10/S11 surfaces mature. Land S15/S16 core after S10's policy-reversal is settled only if sharing the same component-state blob would otherwise churn the version tag twice; otherwise they are independent.

Ingest through the bank — the extension owns ingest (decided "option 1", 2026-07-26; PLAN.md S8)

Decided: loading a sample into the sampler is ONE gesture — capture/import-into-bank AND auto-assign to the active sampler instance — and the extension owns it. The instrument stays a read-only bank consumer; it never captures and never imports. The extension is the right owner: it has arrange access, Media-Explorer access, and the drop-target surface on its own docked panels. Ingest lives in the extension codebase (actions + bank_panel + capture/import add-path), routing through the existing capture add-path and the live "reasampler" seam the instrument already reads.

The three ingest surfaces, with the honest SDK reality (verified against the vendored headers):

• Arrange capture → bank → assign. A one-click action captures the selected item / time-selection into the bank (reusing the existing capture request path — CountSelectedMediaItems / GetSelectedMediaItem + GetSet_LoopTimeRange are already the capture inputs) and assigns the resulting Sample id to the target instance. It never inserts a timeline item — the capture/placement separation is load-bearing; assignment is a bank-index + instance-selection act, not a placement.
• Media Explorer import → bank → assign. The Media-Explorer surface is thin: OpenMediaExplorer (open/select a file) and MediaExplorerGetLastPlayedFileInfo (read the one last-played/selected file path + its selection range/pitch/vol/rate) are the whole contract. There is no enumerate-selected-files and no register-a-drop-handler-on- the-Media-Explorer API. So ME import is single-file, pull-on-action — an action fired while a file is selected in the ME — not a push/drop from inside the ME. Spike: confirm MediaExplorerGetLastPlayedFileInfo returns a usable path+range for a merely-selected (not-yet-played) file, or whether a play is required first.
• Drag-and-drop onto ReaSampler surfaces. REAPER exposes no drag-drop registration API. Drop handling is on ReaSampler's own HWNDs via SWELL/Win32 (WM_DROPFILES / IDropTarget on the docked bank_panel HWND — the surface the panel already owns) → ingest → assign. Assess-and-flag (spike, not promised): a drop onto the VST3 editor window — whether the IPlugView HWND can accept an OS file drop and relay it to the extension as a bank-ingest request (the instrument does not ingest; it forwards a request over an agreed seam). This crosses the two-artifact boundary and the relay is unproven; if gnarly, drop-onto-panel is the shipped path and drop-onto-editor is deferred.

The assign seam. The ingest action names the target instance (lean: the active/ last-focused instance, discovered via the host context the bridge already resolves) and hands it the new sample id — the same instance-owned selection state S4 already persists, so a reload picks it up. With the change-detection seam (below) the assignment refreshes hands-free; without it the ingest action pokes the target instance's reload directly.

Guardrail (load-bearing, restated): ingest is an extension act. Any instrument code path that captures, imports, inserts a timeline item, or writes back into the bank is a bug — the instrument reads and plays only.

Bank-generation change-detection — hands-free refresh (decided 2026-07-26; PLAN.md S9)

Instances reference sample ids. So a recapture (M10) landing under the same id — or an ingest (S8) touching the active bank — should refresh playing instances hands-free, without re-opening each editor. The missing trigger: a bank-generation counter in "reasampler" ext-state.

• Writer (extension). A monotonic bank-generation counter, stamped into "reasampler" ext-state under a new forever-stable ext_keys.h constant, bumped on every bank-content mutation that changes what an instance would play (capture add, recapture-in- place, sample-remove, move/copy affecting the active bank). Additive to the persist blob; defaults to 0 for projects saved before the stamp exists.
• Reader (instrument). Poll the generation over the bridge on a safe off-audio-thread cadence (a UI/timer tick, never process), compare to the last-seen value, and call the existing off-thread reloadFromBank() on change — reusing S4's atomic pointer-swap handoff (graveyard-reclaim) so a mid-play refresh does not glitch. No new audio-thread work; no allocation in process.
• Cadence + safety. A low-frequency UI timer, coalescing multiple bumps between polls into one reload (build-time residual). The read already tolerates a stale value by design (it reloads on the next poll). Must-verify before build: no torn-read hazard on the single integer generation key for a bridge read on the instrument's UI/timer thread concurrent with an extension write.

This seam serves both S8 ingest and M10 recapture; the writer side is extension-only and independent of S8, so it can land alongside either.

Design-system foundation — moved to Phase L (2026-07-26)

The visual design language moved out of Phase S into its own Phase L. The design-system content that stood here (the toolkit assessment, the shared LICE drawing kit S0-DS, and the dock-panel refresh S14) has been lifted into Phase L (Look-and-feel) on dev, taken up by a parallel team so Phase S feature work proceeds ungated. S0-DS is now Phase L point L1 (the shared kit); S14 is now L2 — and, per Daniel's DS-3 call, expanded from a light re-skin into a thorough dock-panel layout redesign that lays out the full M11-aware button inventory before applying the kit; the VST editor + embed-strip restyle is now the explicit L3 point (gated on Phase S landing on dev). The design-system forks DS-1 (LICE + WDL free game, no external frameworks), DS-2 (Direction B "Neon Console" + Direction C's spectral keyboard strip), and DS-3 (thorough panel layout redesign) are all SETTLED (Daniel, 2026-07-26).

Authoritative from here: PLAN.md §Phase L + CONTEXT.md §Phase L on dev, and docs/product/visual-design-language.md (on dev). Phase S's S10–S13 build their interaction UX with the current drawing and adopt the Phase L kit when it lands — they are not gated on Phase L. LICE/SWELL-only, the pure-geometry-module discipline, RT discipline, and the read-only-over-bank / VST3-class-UID-unchanged guardrails all hold exactly as before — a visual refresh is not a data-ownership or compat event.

ReaSampler 9000 — the UX overhaul (S10–S13; DAW-tested S1–S6, "the UX is awful")

The bar is set: better than ReaSamplOMatic5000. Daniel DAW-tested the S1–S6 instrument and the verdict was that it works but the UX is unacceptable — "this is supposed to be better than ReaSamplOMatic5000." The S1–S6 editor was a spike-grade LICE panel: a clickable sample list, zone rows each carrying seven tiny ±1 nudge/delete mini-buttons (low-/low+/high-/high+/root-/root+/delete), text-only labels, no keyboard visualization, no waveform, no drag interaction of any kind, no scrolling for long lists. Setting a zone from C1 to C4 by clicking "+" thirty-six times is the catastrophe; the rest (no way to see a sample, no loop editing by eye, unreachable rows past the panel bottom, a fixed envelope) compound it. The overhaul is scoped as S10–S13, sequenced so the friction Daniel feels every test pass is removed first.

Workflow hierarchy (REVISED 2026-07-26 — Daniel; supersedes the keymap-first S10)

The overhaul is reframed around the actual workflow, not a keymap. Daniel's directive, distilled: a giant list of "item" blocks is visually useless; optimize for working with individual captures, not a huge list of everything. The settled hierarchy:

1. Primary flow = one capture, fast. Most instances play a single capture. The metric is time-to-first-note: open → pick a capture → see it (waveform/peaks) → play it. The default editor face serves this, not a zone table.
2. Fresh instance is SILENT — nothing auto-selected (policy reversal of S4). On open with no stored selection, the instrument plays nothing and shows a clear empty state ("pick a capture") — it does not auto-play sample #1. This deliberately reverses the S4 "first sample plays" convenience: the selectSample first-sample fallback and the processor's Tier-0 fallback that resolved it are removed; an empty stored id resolves to silence. (Recorded as a reversal, not a regression.)
3. Capture browser, not an item list. Scannable cards/rows with peak thumbnails (the Sample peaks bank_model already carries — the same data the dock panel thumbnails draw), name, and a root/key badge where present; filterable by bank (bank_book named banks). A "giant list of item blocks" is the anti-pattern — the browser is designed for scanning by eye.
4. Graphic, descriptive controls with a guided fast path. Once a capture is picked, a prominent, self-explanatory single-capture setup surface (root note, play-mode basics, level). The keyboard strip serves the single-capture case first (shows where the capture sits / its root); drag matters most when zoning.
5. Zones demoted to secondary (nice-to-have). Multi-zone keymap editing becomes an opt-in "Zones" panel (S10-Z), not the default face — "most of the time the zones won't be used." The keyboard-strip drag machinery is still built, but in service of the capture-first layout.

What "better than RS5K" means, specifically (not vibes). RS5K's genuine strengths — match or beat each: (1) drag a file straight onto it loads the sample (our S13 relay); (2) note-start / note-end range with a visual sense of the keyboard (our S10 keyboard strip — RS5K's own range UI is two number fields, so a draggable strip beats it); (3) a waveform with draggable start/end/loop markers (our S11); (4) ADSR sliders (our S12); (5) velocity layers / round-robin (Tier 2 — held, not in this overhaul). RS5K's real weaknesses are our opening: its one-sample-per-instance model forces track sprawl (one RS5K per drum) and it has no multi-zone view in a single instance — ReaSampler 9000 is multi-zone in one instrument by design (S5), so the opt-in Zones panel showing all zones at once is a capability RS5K structurally lacks. But per the reframe, the default face is the single-capture fast path (browser + setup), and multi-zone is the demoted nice-to-have. "Better than RS5K" = a fast single-capture browser where RS5K makes you drag a file blind, direct-manipulation where RS5K uses number fields, multi-zone-when-you-want-it where RS5K is one-shot, and bank-integrated ingest where RS5K is file-at-a-time.

Constraints (unchanged — settled, do not re-open): LICE/SWELL drawing only (no toolkit change — D-A settled); all layout/hit-test math in pure geometry modules (mirror of mode_switch / editor_geometry / embed_strip), the draw + drag-state machine in the shell; RT discipline untouched (every edit commits off the audio thread via the existing commitMapAndReload → off-thread reloadFromBank → atomic swap); the instrument stays a read-only bank consumer (loop/root/ADSR edits are the instrument's performance map, D-B — never written back to the bank); component-state persistence and read-only-over-bank stay settled.

• S10 — capture-first editor: browser + guided single-capture setup (REVISED 2026-07-26). The default face is the capture browser (scannable cards with peak thumbnails from the Sample peaks bank_model carries, name, root/key badge; bank filter over bank_book banks) feeding a guided single-capture setup (root note, play-mode basics, level). Fresh instance is silent, nothing auto-selected — the S4 first-sample fallback is removed (empty stored id → silence + a "pick a capture" empty state). New pure modules: capture_browser (card/grid layout + hit-test) and keyboard_strip (key-span↔pixel via the embed_strip idiom; a root marker for the single loaded capture; pixel→note; drag-delta resolver; per-zone bar rect + edge-grab hit regions for the opt-in Zones panel). Shell extends the click-only wndProc to a WM_MOUSEMOVE/WM_LBUTTONUP drag-state machine with live feedback, one coherent edit on release. Multi-zone keymap editing is an opt-in "Zones" panel (S10-Z), not the default — the demoted nice-to-have; it reuses the same strip geometry + drag machine (edge = resize, body = move, key = root) and retires the seven ±1 nudge buttons per row. Built with the current LICE drawing; adopts the Phase L kit (L1) when it lands (drawn through the shared component kit rather than flat LICE_FillRect/GDI once available) — not gated on Phase L; the drag machine's WM_MOUSEMOVE tracking also lights the kit's hover states at near-zero marginal cost once the kit is present. Boundary shifts (from the reframe): the "sample list" S12 was to scroll/search is now this browser — the card layout, peak thumbnails, and bank filter are S10's; S12 keeps scroll + type-to-filter search layered over S10's browser (bank filter picks the bank, search narrows within it). The waveform S11 makes loop-editable is the same waveform S10 shows read-only for the picked single capture ("see it").
• S11 — waveform view + draggable loop points. Selecting a zone shows its sample's waveform (peaks via the existing peaks module over the shell's already-decoded PCM — no new decode/WAV path) with draggable start/end/loop-start/loop-end markers that snap to zero-crossings (the S2 zero-crossing-aware requirement). A dragged loop is a per-zone loop override (additive on PerformanceZone, same shape as rootOverride; seeded from the S2 bank intrinsic, never written back). Marker/waveform geometry pure (frame↔pixel, marker grab regions, clamp start≤end, zero-crossing snap helper).
• S12 — scale + ergonomics. Scroll (wheel + scrollbar) over S10's capture browser so a bank longer than the panel is fully reachable, and a type-to-filter search that narrows the cards by name, composing with S10's bank filter (bank filter selects the bank; search narrows within it). *(Boundary shift from the 2026-07-26 reframe: the browser card layout, peak thumbnails, and bank filter are now S10's; S12 = scroll
  • search layered over that browser.)* Direct numeric entry for zone low/high/root (a click-to-type field over the strip, for precision the drag can't hit — Zones-panel-scoped). An ADSR editor — four draggable controls over the S3 AdsrParams (the math already exists and is wired into the voice engine; today the envelope is a fixed default). Scroll/search/slider/entry layout pure; ADSR + (implicitly) any exposed parameters become per-instance component state (additive, version-bumped, back-compat).
• S13 — drop-to-load (the S8 relay, in the editor). Dropping an OS file / media item onto the editor window ingests into the bank + assigns to this instance — the RS5K "drop a file straight on it" affordance. The instrument does not ingest: the editor's drop handler relays a bank-ingest request to the extension (S8's option 1), which performs the capture/import + assign; refresh is hands-free via S9 (or a direct reload without it). Cross-artifact relay is the S8-flagged spike — proven-and-shipped or degrade to the docked-bank_panel drop path with a clear affordance. Never inserts a timeline item (capture/placement separation intact).

Sequencing (recommendation, argued below in this section's tail). S10 first — under the reframe it now carries the whole felt win: the empty-state / no-auto-select fix, the capture browser (peak thumbnails, bank filter) that replaces the useless item list, and the guided single-capture setup that retires the nudge buttons. This is the entire "the UX is awful" wound, and time-to-first-note is the metric it moves. S11 (waveform + loop) and S12 (scroll/search over the browser, numeric entry, ADSR) follow — both lean on S10's browser + drag machine, and S11's waveform is the same surface S10 shows for the picked capture. S13 depends on S8's ingest seam, so it sequences after S8. Against the queued engine work: S10 should land before or interleaved with S7 (stereo). S7 is a real engine capability (stereo capture in true stereo) and touches the DSP Daniel smoke-tests — but the reason he'll keep smoke-testing is the editor, and today every test pass is taxed by the nudge-button UX. Fixing what he feels first (S10) makes every subsequent S7 test less painful; there is no hard dependency either way (S7 is engine/bus, S10 is editor/geometry — orthogonal). Honest counter: if the stereo sound is the thing blocking real use, S7 first is defensible — but "it works, the UX is awful" points at the editor as the live wound, so S10 leads.

Product name — ReaSampler 9000 (Daniel, 2026-07-26)

The MIDI-playback instrument's product name is ReaSampler 9000. The extension stays ReaSampler (capture + organization); the instrument is ReaSampler 9000 (playback). Set by Daniel on DAW-testing the S1–S6 instrument, alongside the UX-overhaul directive.

• Propagate the display name across user-visible surfaces: the VST3 class display name string in the factory registration, the factory vendor/name strings, the IPlugView editor title band (currently "ReaSampler Instrument"), the S6 embed-strip label, and the Phase S docs.
• Do NOT change the VST3 class UID. Instances in already-saved projects key off the class UID; changing it orphans every existing instance in every saved project. The UID is a forever-stable contract (mirror of the command-id / ext-state-namespace forever-stable strings).
• S-NAME-1 SETTLED (Daniel, 2026-07-26): rename the binary filename too. The on-disk module name is renamed to match the product (e.g. reasampler_9000.vst3), not just the display strings. Full rename surface: CMake OUTPUT_NAME on the second VST3 target, the factory vendor/name strings, the editor title, and the embed label. The class UID stays locked as the compat anchor.
• Compat verification (must-DAW-verify before shipping the rename). The working assumption is that REAPER rebinds a saved instance by its VST3 class UID, not by the module filename — so a filename rename with an unchanged UID keeps saved projects working. This is a to-verify assumption, not a confirmed fact: a web check surfaced a JUCE/VST3-replace-VST2 case suggesting REAPER's binding can be more nuanced than "UID only" (an FXID match is involved), so it is not safe to assert UID-only rebinding from source. DAW-verify: save a project with an instance under the old filename, rename the module, reopen, and confirm the instance rebinds and restores its state. If REAPER keys partly on filename, fall back to keeping the current filename (display-strings-only) and record that as the shipped choice.

REAPER / Steinberg API surface (verify all signatures)

• VST3 SDK (a new vendored dependency — vendor it at the spike). FUnknown and the IComponent / IAudioProcessor / IEditController interface family; the SingleComponentEffect / EditControllerEx1 / AudioEffect base classes; the class factory (GetPluginFactory + factory macros); IPlugView for the editor; ProcessData / ProcessSetup for the hot path. Verify interface members, the base-class overrides, factory-macro spellings, and the Windows module-export symbol names (InitDll/ExitDll/GetPluginFactory) against the vendored SDK at the spike — the framing doc flags several of these as experienced estimates.
• REAPER VST-host bridge. hostcb opcode 0xdeadf00d (resolve API function by name) and 0xdeadf00e (host context); the by-name resolution of GetProjExtState/SetProjExtState/EnumProjExtState. Verify against vendor/reaper-sdk/sdk/reaper_plugin.h + video_processor.h + reaper_plugin_functions.h.
• Embedded UI (D-D, later point). IReaperUIEmbedInterface and the embed message/lifecycle contract — verify against vendor/reaper-sdk/sdk/reaper_plugin_fx_embed.h before use.
• VST3 bus arrangement (S7 channel mode). setBusArrangements / getBusArrangement and REAPER's mono/stereo instrument-bus expectations — verify against the vendored Steinberg SDK + reaper_vst3_interfaces.h.
• Ingest surfaces (S8). InsertMedia is the placement path (untouched by ingest); CountSelectedMediaItems / GetSelectedMediaItem + GetSet_LoopTimeRange are the arrange-capture inputs (already the capture path's); OpenMediaExplorer + MediaExplorerGetLastPlayedFileInfo are the whole Media-Explorer contract (thin — no enumerate-selected, no ME-drop-handler). Drop handling is SWELL/Win32 on ReaSampler's own panel HWNDs — REAPER exposes no drag-drop registration API. All verified against reaper_plugin_functions.h.
• Bank-generation seam (S9). New forever-stable ext_keys.h key for the generation counter; read over the same bridge GetProjExtState path S4 already uses. No new API — confirm no torn-read hazard on the integer key.
• WDL pitch/resample (S15/S16). Verified this pass: vendor/WDL/WDL/resample.h (WDL_Resampler — sinc/linear, RT-suitable) and vendor/WDL/WDL/simple_pitchshift.h (WDL_SimplePitchShifter — time-domain OLA, set_formant_shift empty stub) are the whole pitch/resample surface; no elastique / formant-preserving / time-stretch in the tree. S16 modulation is hand-rolled (no WDL). If the held sinc-repitch quality upgrade is taken, verify WDL_Resampler streaming/prealloc against the per-voice RT budget before use.
• LICE/SWELL editor. Reuses the bank_panel LICE/SWELL drawing surface; verify the IPlugView↔LICE window/bitmap bridge at the spike (window creation, sizing, event routing) — the least-trodden edge of the phase.
• LICE design-kit surfaces — moved to Phase L. The shared LICE drawing-kit surface verification (LICE_GradRect/LICE_RoundRect/AA lines/circles/beziers/polygons + the LICE_CachedFont/LICE_IFont font engine, and the vwnd drawing-craft references) now lives with Phase L point L1 on dev — see CONTEXT.md §Phase L "LICE / WDL API surface". Phase S surfaces (S10–S13) adopt that kit when it lands; they are not gated on it.

Drop-and-load — drag a capture onto a track's FX button (S17 spec)

The gesture. While a capture is dragged out of the bank_panel, a track's TCP FX button becomes a drop zone. Dropping the capture there instantiates a ReaSampler 9000 on that track with the dragged capture already loaded and selected for playback — one gesture from bank to playable instrument. This is the third integration gesture: capture (extension), placement-into-arrange (extension), and now placement-of-the-player (this wave). It is drop-and-load, not drop-to-arrange — no media item touches the timeline.

Why it needs a new drag mode (the CF_HDROP path can't carry it). Today's drag-out (M11) becomes an OS file drag (CF_HDROP via drag_out + drag_out_win) the moment the pointer leaves the panel client rect. REAPER's TCP FX button is not a native drop target that instantiates a plugin-with-a-file, so this feature cannot ride the OS-drag path: an OS drop of a WAV onto the FX area does not create "an instrument preloaded with that WAV." It requires an internal drag where the extension itself tracks the pointer over REAPER's own UI, detects the FX-button hover, and on release drives the insert itself. The extension is the actor for the whole gesture.

The two-part mechanism.

1. Internal-drag hover detection (extension-side, pure + shell). The drag_out pure module gains a third DragGesture beyond Internal (bank-to-bank) and OsDrag (M11) — InstrumentDrop. The gesture decision is refined: leaving the panel client rect no longer immediately means OS-bound. Instead:

   • Pointer inside the panel client rect → Internal (unchanged bank-to-bank drag).
   • Pointer outside the panel but still over REAPER's own window/UI → InstrumentDrop (new — the shell hover-tracks the TCP FX button and highlights it).
   • Pointer left REAPER entirely (Explorer / another app) → OsDrag (unchanged M11).

   The pure module stays REAPER-free: it decides InstrumentDrop vs. OsDrag from position plus an "over-REAPER's-own-UI" predicate the shell supplies (the shell owns the REAPER window/hit query; the pure layer owns the set/boundary algebra). Mirror of how M11 kept decideGesture pure over a rect the shell supplied. The shell then resolves the pointer to a track + FX-button hotspot, highlights it, and on release drives the drop.

2. FX-button drop → add-VST + load-capture (extension-side shell, then instrument seam). On release over an FX button the shell:

   • Adds a fresh instance: TrackFX_AddByName(track, "VST3:ReaSampler 9000", /*recFX*/ false, /*instantiate*/ <negative>). Verified present in reaper_plugin_functions.h: int TrackFX_AddByName(MediaTrack* track, const char* fxname, bool recFX, int instantiate) — a negative instantiate always creates a new effect (per the header comment); the "VST3:" prefix selects the format. Captures the returned FX index (or -1 on failure).
   • Loads the dragged capture into that instance via the load-capture seam (below).
   • Wraps the whole thing in one REAPER undo point (Undo_BeginBlock2/EndBlock2) so the gesture is one Ctrl-Z — the same discipline the bank verbs use.

The ReaSampler 9000 load-capture seam (the hard coupling — MUST be added; does not yet exist). The Phase S spec today gives the instrument a live-state read seam (it reads bank index + mapping from "reasampler" ext-state via the bridge — §The two seams) but no entry point for an external actor to say "this fresh instance should play this specific capture." Reading the bank is not the same as being pointed at one sample. This wave is the reason to add that seam, and the seam lands inside the instrument (the phase-s artifact), not the extension. Two candidate mechanisms — the choice is a Phase S open question:

• (A) Fresh-instance ext-state handshake. The extension writes a small "pending load" hint into "reasampler" ext-state keyed to the target track/FX (e.g. the capture id + a target GUID); a freshly-instantiated ReaSampler 9000 reads it on init via the same bridge it already uses, claims + clears the hint, and self-selects that capture. Pro: reuses the existing bridge; no new VST3 surface. Con: a cross-process handshake with a claim/clear race to get right; "which instance claims which hint" needs a stable key.

• (B) VST3 setState / preset injection. The extension builds the instance's component state (the same blob getState/setState round-trips) with the capture pre-selected and injects it right after TrackFX_AddByName. Pro: deterministic, no shared-state race, uses the instrument's own persistence format. Con: the extension must know and construct the instrument's state blob format — a tighter cross-artifact coupling to a format that is itself still being built in Phase S; verify whether the extension can set an added FX's state through the REAPER API (candidate: TrackFX_SetNamedConfigParm / a state-set path — not yet confirmed against reaper_plugin_functions.h).

  Lean: (A) keeps the artifacts loosely coupled through the one ext-state seam they already share and avoids the extension hard-coding the instrument's state format — but it inherits the claim/clear race. Daniel to decide (open question below).

Coexistence with the OS drag-out (disambiguation contract). The two OS-vs-internal modes are disambiguated by pointer location, not a mode toggle — the user never picks "OS drag" vs. "instrument drop"; the extension infers it from where the pointer is when released. The M11 boundary (left the client rect) is refined, not replaced: leaving the rect now asks "over REAPER's UI → InstrumentDrop, else → OsDrag." Both M11 OS drag-out and the internal bank-to-bank drag must remain byte-for-byte unchanged in their own regions — this wave only inserts a new middle case. Multi-capture payloads are a disambiguation input too (see open question — instrument drop is naturally single-capture; a multi-capture drag over the FX button is either rejected or loads the first).

Precision / invariant implications (drop-and-load).

• Explicit user-driven placement — consistent with capture↔placement separation. This is a deliberate placement gesture: the user chooses to put a playing instrument on a track, exactly as inserting an item into the arrange is a deliberate act. It does not auto-capture (the file already exists in the bank) and does not insert a media item into the timeline. It instantiates a reader of the bank on a track and points it at one already-captured sample. Capture, placement, and playback stay three distinct acts; this is placement-of-the-player, not a capture and not a timeline insert.
• No private sample copy. The instantiated instrument consumes the one authoritative bank (it resolves the WAV via the shared M4 project-relative machinery like any ReaSampler 9000 instance); the seam hands it a reference (a capture identity), never a copied file. Any path that copies bytes into the instance is a bug.
• The internal drag stays pure-decidable and testable. The new InstrumentDrop gesture is decided in the drag_out pure module (REAPER-free) over a shell-supplied predicate; the M11 drag_out unit tests must not regress.

Open questions (Daniel / Phase S team to decide).

• Load-capture seam mechanism (A vs. B above). The single load-bearing Phase S design choice this wave forces. Lean is (A) for loose coupling; needs Daniel's call.
• Multi-capture drag over an FX button — reject (only single-capture drags arm InstrumentDrop), or load the first / a keymap of all? Tier-0 leans reject-or-first; a multi-capture keymap load is a Tier-1 stretch.
• FX-button hotspot vs. whole TCP. Does the drop zone have to be the FX button specifically, or is dropping anywhere on the target track's TCP enough (simpler hit resolution, arguably clearer target)? Depends on what the SDK exposes (see must-verify).
• Numbering vs. the phase-s worktree (reconciled on merge to dev). Authored on dev as a provisional S7; renumbered to S17 on merge, since the worktree's authoritative Phase S set (S7 stereo, S8 ingest, S9 change-detection, S10–S16) took the lower labels. Drop-and-load stays a distinct integration gesture (drop onto a track's FX button) — a sibling of but not the same as S8 (ingest into the bank) and S13 (drop-to-load inside the editor). See PLAN.md §S17.

Must-verify before build (drop-and-load).

• TrackFX_AddByName — verified present (reaper_plugin_functions.h): signature and the "VST3:"-prefix + negative-instantiate semantics confirmed from the header.
• Pointer→track / FX-button hit resolution during a drag — not yet confirmed. Candidates: GetTrackFromPoint / GetThingFromPoint (verify names + signatures against reaper_plugin_functions.h); whether the FX button specifically is addressable vs. the TCP as a whole is an open verification that also decides the "hotspot vs. whole TCP" question.
• Instance state injection (only if seam mechanism (B) is chosen) — whether the extension can set a just-added FX's state via the REAPER API (TrackFX_SetNamedConfigParm or similar) — not yet confirmed against reaper_plugin_functions.h. Moot if (A) is chosen.

Non-goals / guardrails

• The instrument never captures and never inserts into the arrange. Playback is a read-only act over the bank. Any instrument path that captures, places a timeline item, or writes back into the bank is a bug — reject in review.
• The instrument keeps no private copy of the samples. It consumes the one authoritative bank; per-instance sample stores are a non-goal (they refork the source the one-source-multiple-views instinct keeps single).
• The instrument never ingests (S8). Capture, import, and drop-ingest are extension acts; the instrument only reads and plays. A drop onto the editor window (if the spike proves it viable) is relayed to the extension as an ingest request — the instrument never writes the bank itself.
• Channel mode is a performance choice, not a bank fact (S7). The mono/stereo toggle is per-instance component state, never written to Sample or the bank (D-B). The bank's per-sample channel-count intrinsic is a file fact; the play mode is the instrument's.
• No cross-platform / multi-format. Windows-only, VST3-only, REAPER-only (D5). Do not add an AU/AAX/VST2/CLAP wrapper, a mac/Linux build, or a standalone host target.
• The pure core stays REAPER-free and VST3-free. The voice engine / envelope / keymap / repitch module takes no VST3 or REAPER type at its boundary — the shell marshals. Any VST3 or REAPER type leaking into the core is a bug (the D3 split).
• Additive to the extension. The Sample intrinsic-field addition is additive (new optional fields; no existing field or BankIndex behavior changes); everything else in Phase S lives in the second artifact and does not alter the extension's M/D/B/R/V pillars.
• Do not spec Tier 2/3. Tier 2 is held (noted, not specified); Tier 3 is optional-forever. Do not let their feature lists drive Tier 0–1's build shape. Note: S7 stereo is not a Tier-2 feature — it is a channel-count dimension on the existing Tier 0–1 engine, orthogonal to Tier 2's velocity-layers / round-robin / per-sample trim. (S7's stereo loop read is the same loop the core already has, extended per-channel — not the Tier-2 "sustain loops" feature.) Likewise S15/S16 (Trigger/Gate modes + pitch envelope) are Daniel-directed engine features on the Tier 0–1 core, not Tier 2/3 — the AHDSR hold, Trigger one-shot, start point, and AD pitch envelope are orthogonal amplitude- shape / read-rate dimensions, not the held velocity-layers / round-robin / filter work.
• S15/S16 params are performance choices, not bank facts. Play mode, start point, %-length, fades, AHDSR, and the pitch envelope are per-instance performance-map state (component state), never written to Sample or the bank (D-B). The bank carries file facts (root note, loop intrinsic, channel count); the instrument owns how they are played.
• S15/S16 stay channel-count-agnostic (S7 interplay). The mode/envelope logic is per-frame amplitude and read-rate, independent of the S7 channel dimension. Any S15/S16 code that assumes a fixed channel count (mono) — rather than operating per-frame pre-mix — is a bug that would collide with S7. Spec and build them channel-agnostic.
• Trigger ignores note-off; choke is out of scope (S15). In Trigger mode note-off is a no-op and the one-shot plays to playEnd. Choke-on-note-off / choke-groups are held (fork S15-F1, Tier-3-adjacent) — do not add a choke path in S15.
• Pitch envelope is off by default (S16). Default-disabled → offset always 0 → ratio_ unchanged → playback bit-identical to pre-S16. A regression that applies pitch modulation when the envelope is off is a bug.
• Drop-and-load must not regress the two existing drags. S17 inserts a new middle case (InstrumentDrop) between the M11 OS drag-out and the internal bank-to-bank drag; both existing gestures stay byte-for-byte unchanged in their own regions. Drop-and-load never inserts a media item into the arrange and never copies sample bytes into the instance — it hands the new instance a reference to an already-captured bank sample.
• Verify Steinberg SDK, bridge, embed, and LICE-view surfaces against the vendored headers before use — several §1a claims are experienced estimates until the spike confirms them.

---

Look-and-feel — visual design language (Phase L)

New pillar, own lettered phase, taken up by a parallel team. Phase L is the whole-system look-and-feel effort — a shared LICE drawing kit and the surfaces that adopt it — that replaces the flat "temple os" drawing (opaque LICE_FillRect blocks + raw GDI DrawTextA) with a modern, sleek 2026 dark synth look across ReaSampler (the extension's docked bank panel) and ReaSampler 9000 (the Phase S VST editor + embed strip). Namespaced L (Look-and-feel), orthogonal to and ungated by the M/D/B/R/V/S pillars. Product framing, the settled decision record (DS-1/DS-2/DS-3 all SETTLED 2026-07-26), palette, the three visual directions, and the toolkit assessment: docs/product/visual-design-language.md. Same standing discipline: verify every LICE/ WDL/SWELL API name/signature against vendor/WDL before use. When a point lands, doc-keeper moves it to COMPLETED.md.

What it is

A shared LICE-based drawing kit (palette + type scale + component-draw layer) and the surfaces that consume it. The kit is the one source of drawing for the whole system — a button, row, slider, or waveform looks identical in the bank panel, the embed strip, and the VST editor because it is the same kit function (the one-source-multiple-views instinct, applied to drawing). The current "temple os" look is the system drawing at the floor of LICE (flat fills, GDI text, no gradients/AA/rounded/hover); the kit lifts every surface to LICE's actual ceiling — which REAPER's own themed UI and SWS prove is a modern dark UI.

Settled decisions (Daniel, 2026-07-26 — reasoning in docs/product/visual-design-language.md §6)

• DS-1 — toolkit: LICE + WDL free game, no external frameworks. Draw with LICE directly (gradients via LICE_GradRect, AA rounded via LICE_RoundRect/LICE_Line, cached AA text via LICE_CachedFont). Reuse any useful WDL/vwnd piece — skin/image helpers, draw idioms, a specific control (e.g. virtwnd-listbox for a long scroll list) — where it beats re-deriving; "don't reinvent the wheel." Reject external frameworks (iPlug2 / JUCE / VSTGUI — they re-open the settled bare-SDK+LICE build shape to solve a look problem that is not a toolkit-ceiling problem). Caution, not a ban: keep hit-test geometry in pure CTest-covered modules — do not import vwnd's retained-mode object model wholesale (its controls own their hit-test internally, which would move geometry into untestable shell code and undercut the pure/shell split).
• DS-2 — visual direction: Direction B ("Neon Console") + Direction C's spectral keyboard strip. SETTLED 2026-07-26, REVISED 2026-07-26 (Daniel) — three-accent pastel accents AND REAPER-grey neutrals (palette-only). Neutral surfaces (revised): the neutral ladder moved from near-black up into REAPER's mid-grey theme family so the dock reads as part of REAPER, not a black slab — bg/base ≈ #2b2b2b (REAPER chrome), bg/panel ≈ #333333, bg/cell ≈ #3a3a3a (REAPER track bg), line/hairline ≈ #4a4a4a, text/primary ≈ #dcdcdc, text/dim ≈ ~#a0a0a0+. Elevation-ladder discipline unchanged (base < panel < cell by a few %, micro-gradient + inner highlight/shadow carry elevation, not hard borders). Accent layer (revised): changed from one electric cyan to three pastel accents: accent/primary = pastel lime green (~176,224,152, the live/active/selected signal), accent/secondary = pastel teal (~132,214,208, categorical role A), accent/tertiary = pastel purple (~194,170,232, categorical role B); accent/hot is a lighter pastel-lime tint (~200,236,178). These are starting values — the implementer locks final hex against the theme module's WCAG tests, staying within the pastel intent ("soft-side" floor: the most pastel value that still clears, not re-saturated toward neon). The spectral (hue-mapped) keyboard strip — signature surface, glow-on-active as a static drawn state — is now a pastel sweep anchored on the three accents (pastel-lime low → pastel-teal mid → pastel-purple high). A stylish/bundled font upgrade was considered and DECLINED (Daniel, 2026-07-26) — no font bundling or redistribution; the kit keeps its current cached-font face and no new typeface is specified (§3.1). The kit palette stays abstract (role→color, one constants block — now three accent roles), so the direction is a single-file change. Tight pairs flagged for the implementer (re-verify against the GREY ladder, not near-black): text/dim on grey (mid-grey-on-mid-grey — lands ~4:1 at start, must be lifted toward ~#a8a8a8+ to clear AA 4.5:1 body on #333/#3a3a3a, the classic floor failure); the three pastels as state indicators / active fills on bg/cell (the pastel cushion shrank from ~15:1 on near-black to ~6:1–7:1 on grey — still clears 3:1 but must be re-checked; if any drops below floor, nudge that hue slightly deeper within the pastel intent); body text on a pastel fill (light-grey on a light pastel can drop below AA 4.5:1 — use dark labels on pastel fills or darken the fill); and secondary vs. tertiary distinguishability (both cool/desaturated — verify they read as distinct categories). Full reasoning: docs/product/visual-design-language.md §2.1 + §4 + §6.
• DS-3 — dock-panel scope: a thorough layout redesign, not a light re-skin. L2 lays out the full button/affordance inventory — including M11's action-trigger buttons + keybinding-help labels — intuitively, uncluttered, and useful, then applies the kit. Sequenced after M11 merges so it designs against the actual landed button set.

Palette + the "punch" rule (Daniel's standing taste)

Dark, modern, visual punch over conservative contrast — now pastel, on REAPER-grey. The palette is defined by role, not hardcoded hue: bg/base, bg/panel, bg/cell, line/hairline, text/primary, text/dim, accent/primary, accent/secondary, accent/tertiary, accent/hot, warn. Neutral surfaces (DS-2 revised, 2026-07-26): the neutral ladder sits in REAPER's mid-grey theme family, not near-black, so the dock reads as part of REAPER — bg/base ≈ #2b2b2b, bg/panel ≈ #333333, bg/cell ≈ #3a3a3a, line/hairline ≈ #4a4a4a, text/primary ≈ #dcdcdc, text/dim ≈ ~#a0a0a0+. A modern dark UI is built from elevation layers, not borders — surfaces gain a micro-gradient (LICE_GradRect, a few percent lighter at the top) + a 1px inner top-highlight / bottom-shadow (the vwnd trick) instead of flat fills; the grey ladder keeps this discipline (base < panel < cell by a few %, subtle steps like REAPER, not hard outlines). The three accents carry categorical meaning: primary (pastel lime) = live/active/selected, secondary (pastel teal) + tertiary (pastel purple) = supporting categorical distinctions (kinds, not intensity — primary is always "what's live now"). WCAG-floor discipline (pastel + grey re-read): moving neutrals up into grey makes two pairs harder — (1) text/dim on bg/panel/bg/cell is the classic mid-grey-on-mid-grey floor failure (must be lifted light enough to clear AA 4.5:1 body on the greyest surface it draws on), and (2) the three pastels as state indicators on grey have a shrunken contrast cushion (~6:1–7:1 vs. ~15:1 on near-black — still clear 3:1 but re-verify; nudge a hue slightly deeper within the pastel intent if it drops below floor). Text-on-a-pastel-fill (AA 4.5:1) remains a tight pair. Take the most pastel value that still clears the floor on grey — approached from the soft side, never re-saturated toward neon "to be safe," except where the grey floor forces a small deepening. The warn role (red/amber) is reserved only for byte-deleting or clip states (prune, delete). Because the palette is role-based in one constants block, the settled-and-revised B+pastel+grey+spectral direction (DS-2) is one file.

"Speed is the selling point" — a design constraint, not a tagline

The UI must feel instant, and no decoration may cost that. Sub-frame hover/press/drag feedback repainted immediately on the input message (instant acknowledgment is the perception of speed); zero-jank via the preserved double-buffer discipline (draw to LICE_SysBitmap, single BitBlt); region-scoped InvalidateRect during a drag (build-time residual). No decorative animation — no tweens/fades/pulses; the only permitted motion is a level/meter readout following the audio directly (as the embed strip already does). Direction C's glow/bloom is a static drawn state, never a pulse. The "fast" feeling is typography + hover + no-jank, not motion.

Kit architecture (the pure/shell split)

Pure (no LICE, no REAPER types, unit-tested — the mirror of mode_switch/bank_grid):

• theme/palette module — role→color mapping, direction-selectable via one constants block (the B + REAPER-grey-neutral + three-accent-pastel + pastel-spectral values). Pure; unit-tested that each text-on-surface pair clears its WCAG floor (the "punch" rule made testable). DS-2 revision (2026-07-26): (a) the neutral ladder moves from near-black up into REAPER's mid-grey family (bg/base #2b2b2b / bg/panel #333333 / bg/cell #3a3a3a / line/hairline #4a4a4a / text/primary #dcdcdc / text/dim ~#a0a0a0+), and (b) the accent role expands from one to three (accent/primary lime, accent/secondary teal, accent/tertiary purple) with the spectral ramp a pastel sweep anchored on those three — all still confined to the one constants block + the shell's font lifecycle. The WCAG test must be re-run against the GREY ladder — it gains the newly-tight grey pairs: text/dim-on-grey (AA 4.5:1) and each pastel-as-state-indicator on bg/cell (3:1), plus the existing text-on-pastel-fill pairs. (No font change: a bundled-font upgrade was declined — the cached-font set keeps the kit's current face.)
• Component geometry/hit-test helpers — button rect, slider track/handle geometry, list-row rect + hover hit-test, and any new layout module L2 needs (an action-bar layout module). No LICE, no host types; CTest-covered. Existing pure modules (bank_grid/tab_strip/mode_switch) stay the source of truth for what they own.

Shell (LICE-facing, DAW-verified — thin draw layer):

• Draw kit — fillSurface (micro-gradient + inner highlight/shadow), drawButton/drawSlider/drawListRow/drawWaveform/segmented-switch/tab draw, and a shared text() over a cached LICE_CachedFont set (title/label/value-mono/micro). Owns the cached-font lifecycle. Honors the interaction state model (rest/hover/active/pressed/ dragging/focus/disabled). DS-1: WDL/vwnd reuse is assessed here at build time — reuse a vwnd piece where genuinely cheaper, else draw on LICE; hit-test geometry stays pure regardless.

L2 dock-panel layout contract (the M11-aware inventory)

DS-3 makes L2 a layout design, not a skin pass, because M11 adds a real button inventory. L2 must place every affordance below without crowding the grid (the centerpiece), grouping by task:

Existing (landed / specced):

• Bank grid — thumbnails, multi-select, keyboard nav, audition, focus ring.
• Design View segmented mode switch ([ Arrange | Design ]) + per-mode membership count.
• Multi-bank: named-banks tab strip (LICE-drawn, overflow/scroll), active-bank indicator, pool/banks full-height toggles, create/rename/delete/activate-bank affordances, per-selection move / copy / remove sample menu.
• Prune button (R-E) — the byte-deleting action; warn-colored, set apart.

M11 adds (dev PLAN.md §M11 — merging to dev now):

• Action-trigger buttons — clickable buttons firing the capture + provenance action family directly (capture item / capture track scopes, re-capture from source, resample-and-mute-source, batch capture, conform-on-insert, insert-at-cursor, drag-out, null-test verify). A cluster.
• Keybinding-help labels — each capture/provenance action surfaces its current key binding (e.g. "Capture Item → F5") or an "unbound"/"—" marker.

Layout mandate: group by task (capture / organize / reclaim / view), not by phase; a compact action bar/toolbar for the frequent capture actions (icon+label, keybinding as a micro sub-label), an overflow/menu for the rare ones, header space for the mode switch + active-bank indicator, the bank tab strip + move/copy/remove organize cluster together, prune set apart and warn-marked. Density is a design decision — 8px grid, elevation layers over hairlines, hover on every interactive element. Then apply the L1 kit to draw it. New layout math goes in a pure geometry module; bank_grid/tab_strip/mode_switch stay the pure source of truth for their own hit-testing. L2 sequences after M11 merges so it inventories the actual landed buttons.

L4 dock-panel button layout (top / bottom / footer re-home)

L4 is a second layout pass over the same bank_panel that re-homes the L2 button inventory around frequency and intent. It ships no new action and changes no capture/placement behavior — every button fires an existing registered action; the "capture ≠ placement" load-bearing principle is untouched (the buttons only fire the split acts, they never fuse them). It is ungated by Phase S (the dock panel is on dev) and independent of L3. L4 sequences AFTER the in-flight DS-2 palette-revision branch merges to dev — both rework bank_panel heavily, and landing L4 concurrently would collide the same file. Drawn through the L1 kit in the DS-2 grey-neutral + three-accent-pastel palette (prune stays warn); no palette or font decisions are re-opened here.

The new three-zone structure:

• Top toolbar = capture + placement + maintenance. The capture cluster (capture item, capture track, batch items, batch razor, capture RT), the placement cluster (insert, insert-conform), and the maintenance cluster (re-capture, cancel-realtime) move from the bottom L2 action bar to a top toolbar — the eye's first landing, matching the acts the tool exists for. Icon+label buttons with the keybinding as a micro sub-label (the M11 keybinding-help convention), drawn through the kit.
• Bottom toolbar = Design View tagging + switching. The space the capture/placement buttons vacate holds the Design View action family as buttons: tag / untag selected tracks for a mode, activate Arrange, activate Design, toggle active mode, show-both. These are the registered Design View actions today; L4 gives them a button home here.
• Footer = narrow Arrange|Design toggle · Tail button · … · Prune. The large top [ Arrange | Design ] segmented toggle shrinks to just-wide-enough-for-its-text and moves into the footer at the left, carrying its per-mode membership count as a compact adjacent label. The Tail affordance is converted from a click-zone to a proper kit button (rest/hover/pressed states; click still cycles the tail setting). Prune stays the byte-deleting action — set apart at the far right, warn-colored — the only file-deleting affordance, kept isolated so no benign toggle sits next to it.

Footer affordance order (left → right): [Arrange|Design] toggle · Tail button · … · Prune (rightmost, set apart, warn). The order reads benign/frequent at the left (view-mode toggle, tail-length control — both "how this panel/capture behaves") → destructive/ rare at the right (Prune, isolated), so a mis-click near the left is cheap and the one destructive control is spatially and chromatically distinct.

Pure/shell discipline (unchanged). All new toolbar-row and footer-strip layout math goes in pure CTest-covered geometry modules — extend/mirror action_bar (the toolbar row layout, now instantiated top and bottom) and mode_switch (the now-narrow, fit-to-text footer toggle geometry + hit-test). bank_grid and tab_strip remain the pure owners of their own surfaces' hit-testing. The L1 kit draws; the geometry stays pure. L4 resolves its re-home against the post-palette-revision bank_panel (build-time inventory once that branch is on dev).

L5 dock-panel button refinements (overflow menu · faces+tooltips · opposite-mode tags · Toggle removal · grouping)

L5 is a third refinement pass over the same bank_panel toolbars L4 built — it re-homes and re-labels buttons for legibility; it ships no new action and changes no capture/placement behavior (every button fires an existing registered action; "capture ≠ placement" is untouched — the buttons only fire the split acts). Ungated by Phase S (the dock panel is on dev) and independent of the L3 gate. Drawn through the L1 kit in the DS-2 grey-neutral + three-accent-pastel palette (prune stays warn); no palette or font decision is re-opened. L5 sequences AFTER L4 (both rework the same bank_panel toolbars).

1. Top-toolbar overflow menu. The less-frequent capture variants — Batch Items, Batch Razor, Capture RT, Cancel RT — leave the visible top bar for a right-anchored "⋯ / More" menu button (kit-drawn button; on click a TrackPopupMenu popup lists all four, each entry firing its existing command id via Main_OnCommand). The frequent acts stay on the bar in cluster order Capture → Maintenance → Placement: Capture Item, Capture Track (Capture); Re-capture (Maintenance); Insert, Insert Conform (Placement). The menu button's rect + hit-test is pure (extend/mirror action_bar / prune_button); the popup + dispatch is shell. The menu entries reuse the same resolveBarCommandId path the bar buttons use, so a keybinding and a menu pick fire identically.

2. Short faces + full-name tooltips (drop the ReaSampler: prefix). Button faces carry the terse ActionBarRow.shortLabel (already the case since L4) — L5 formalizes the face set and adds a hover tooltip showing the FULL action name with the ReaSampler: display prefix stripped. Note the prefix (actionDisplayPrefix()) is baked into the registered gaccel action name, not the button face — so the tooltip derives the full name and strips the prefix for display; the button face never carried the prefix. The keybinding is surfaced in the hover tooltip ("phrase — binding" when bound, bare phrase when unbound) via kbd_getTextFromCmd; the micro sub-row is removed from ActionBarSlot. Tooltip mechanism: custom LICE-kit hover-delay tooltip (tooltip pure module) — owns its own hover timer + LICE overlay draw, stays inside the L1 kit (DS-1 "keep drawing in the kit"). Sourced from the registered action phrase (not kbd_getTextFromCmd); ReaSampler: prefix stripped at draw time; tooltip box width clamped to the client so it never overhangs a narrow dock. The hover-timer threshold + overlay draw are the only DAW-bound pieces; the "which button, what text" decision stays pure.

3. Bottom-toolbar Item/Track × Arrange/Design tag buttons, opposite-mode-only. The current Tag Design / Untag pair is replaced by four buttons — "Item: Arrange", "Item: Design", "Track: Arrange", "Track: Design" — in the Tagging cluster. Both action families already exist in the model + actions layer (research-confirmed): the track family is doTag/doUntag on the track selection (VIEW_TAG_DESIGN and VIEW_TAG_ARRANGE, where Tag→Arrange == untag); the item family is doMoveItems on the item selection (VIEW_MOVE_ITEMS_DESIGN / VIEW_MOVE_ITEMS_ARRANGE, driving planItemRetag + lane minting). So the four buttons are layout + enablement wiring over existing actions — NOT new feature work; no view_mode_model / view / actions change is required.

Enablement rule (precise). Let active = view().activeModeId() (the SAME read the footer [Arrange|Design] toggle uses — one source of truth). A tag button's target mode is the mode in its label (Arrange or Design). A button is live iff target ≠ active; otherwise it is drawn Disabled (kit InteractionState::Disabled, Role::TextDim) and its click is a no-op. Concretely:

• Design active → Item: Arrange and Track: Arrange are live (they send the selection to Arrange); Item: Design and Track: Design are disabled (the selection is already there).
• Arrange active → the reverse: the …: Design buttons are live; the …: Arrange buttons are disabled. The disabled predicate is pure (active mode → per-button live/disabled, unit-tested); the shell reads the active mode once per draw and applies it. Item buttons act on the current media- item selection; Track buttons act on the current track selection — matching the existing action bodies exactly (no selection semantics change).

4. Toggle button removed. VIEW_TOGGLE_MODE leaves the bottom toolbar — the footer's [Arrange|Design] toggle (L4) already covers mode switching. The action stays registered (keybinding-bound, FOREVER-STABLE id unchanged); only its button home is removed.

Fate of Activate-Arrange / Activate-Design / Show-Both (FORK — RESOLVED, Daniel's call).

• VIEW_ACTIVATE_ARRANGE / VIEW_ACTIVATE_DESIGN — dropped from the bottom toolbar (actions stay registered, FOREVER-STABLE ids unchanged). The footer [Arrange|Design] toggle is the single mode-switch affordance; the bottom bar is tagging + Show Both only.
• VIEW_SHOW_BOTH — kept as a set-apart button on the bottom toolbar. It is the cross-mode "pin visible in every mode" escape hatch and is not covered by the footer toggle.
• Landed bottom-toolbar inventory: [ Item: Arrange | Item: Design ] · [ Track: Arrange | Track: Design ] ⟩⟩ [ Show Both ] — Tagging cluster (four opposite-mode buttons) set apart from a lone Show Both. No Toggle, no Activate-Arrange/Design (footer toggle owns switching).

5. Semantic-grouping spacing. L4's kBarSpec is buttonGap=4 / clusterGap=16 (4:1). With the bottom bar's cluster boundary now more meaningful (four tag buttons vs. the Show Both remnant) the grouping should read at a glance — widen the inter-cluster gap: start ≈ clusterGap=24 / buttonGap=4 (6:1), tuned in-DAW. One kBarSpec still serves both toolbars (identical button shape top and bottom); only the gap ratio changes.

Pure/shell discipline (unchanged). All new geometry — the top-bar overflow menu-button rect

• hit-test, the opposite-mode enablement predicate, the wider grouping spacing — goes in pure CTest-covered modules (extend/mirror action_bar / prune_button; add a pure enablement predicate). The TrackPopupMenu popup, the Main_OnCommand dispatch, and the tooltip hover- timer + LICE overlay draw are the only DAW-bound pieces; the L1 kit draws. L5 resolves its menu/tooltip/action-id specifics against the landed L4 bank_panel (build-time confirmation).

L7 capture ordering · card metadata · selection styling (ordering model · overlay · tertiary-border selection)

L7 is the next ungated dock-panel pass over the same bank_panel grid L4–L6 built. Unlike L4/L5 (pure layout over existing actions), L7 (1) is a persisted-model change — it adds an explicit per-sample display order with gap-preserving sparse placement to the persisted bank state and its JSON round-trip. (2) and (3) are draw-only. No new capture/placement behavior; the "capture ≠ placement" principle is untouched. Drawn through the L1 kit in the DS-2 grey-neutral + three-accent-pastel palette; no palette/font decision is re-opened. Ungated by Phase S; independent of the L3 gate. Sequences AFTER L6.

Research baseline (dev, confirmed at spec time). The grid draws in BankIndex insertion order: BankIndex holds std::vector<Sample> samples_ in insertion order (all()), and bank_grid::computeCellRects tiles exactly one contiguous rect per sample left-to-right, top-to-bottom — no explicit ordinal, no sparse slots today. Sample already carries lengthSeconds, lengthBeats, and captureTempo (BPM at capture) but no time-signature field. The selected cell is drawn (bank_panel::drawThumbnail) as InteractionState::Active (accent-fill surface) + an accent/primary border + an inverted waveform (bg/base), with focus as a text/primary double-line ring. Drag today: press on a selected cell arms a drag; crossing a threshold begins it; leaving the client rect hands off to OS drag-out (drag_out::decideGesture); a drop on a tab / the other region moves (or copies on Ctrl) the samples to that bank. There is no same-bank in-grid reorder today.

1. Persisted deterministic order + sparse (gap-preserving) placement

Each bank (and the pool) carries an explicit, persisted display position per sample — the grid no longer derives order from insertion order. Positions are gap-preserving: a sample may occupy a slot that leaves earlier slots empty (an empty first row above an occupied second row is a valid, persisted state).

Where the position data lives (the CLAUDE.md constraint). bank_model / Sample are stated untouched by bank_book's wrapping (no bankId on Sample; a bank is a logical grouping over the shared pool). Display position is a per-bank display concern, so it belongs with the bank's membership, not on Sample — a copy of a sample into two banks can sit at different slots. Recommended carrier: a per-Bank ordered position map in bank_book (sample id → slot), leaving bank_model untouched. The exact carrier (ordered id list with gaps vs. explicit id→slot map) is settled at build; the contract below holds either way.

Ordering contract (deterministic, gap-preserving):

• Deterministic: the grid iterates positions in ascending slot order; ties are impossible (one sample per slot). The order is fully determined by the persisted position data, not by insertion order or hash.
• Insert (new capture): a new capture takes the next free slot after the last occupied slot (append). It never fills an earlier gap automatically — a gap is a user's deliberate layout, not a hole to be plugged.
• Delete / remove / prune: removing a sample leaves its slot empty (does not re-pack) so every other sample keeps its position. Trailing empty tail is trimmed for scroll-extent purposes; interior gaps are preserved (confirmed at build).
• Reorder: the user drags a card to a target slot within its bank; the pure reorder mutator moves that sample's position to the target slot, gap-preserving. Drop-into-empty-slot places there; drop-onto-occupied inserts-before and shifts the tail (matching common file-manager reorder) — SETTLED as the default drop. (The Alt+drop-onto-occupied = REPLACE override is specified below under Drag disambiguation, F3.)

JSON round-trip + migration (load-bearing). serialize/deserialize stay lossless including positions (deserialize(serialize(x)) == x). A pre-L7 project blob has no position data → on load it defaults to current insertion order, densely packed (no gaps), so a project saved before L7 is visually identical on first post-L7 load. This default is the migration; it is one-way (once re-saved, the position data is authoritative). Pure and hard-tested — this is a persisted-model change, tested to the same bar as bank_book's existing round-trip + legacy migration.

Undo. A reorder is one Ctrl-Z — the actions/shell layer wraps the mutation in a batched undo point (Undo_BeginBlock2/EndBlock2), matching every existing bank-verb's undo discipline.

Position-model shape (F2 — SETTLED 2026-07-27: interchangeable substrate, NOT fixed slots). The carrier is the plain gap-preserving interchangeable-slot substrate — a per-bank id→slot map (Daniel: "I don't think I want fixed slots MPC style, but the substrate of interchangeable slots is valuable"). It is not M9-shaped: no slot identities, no numbered/addressable slots that persist independent of their occupant, no slot actions, no MIDI-bindable slot numbers, no capture-to-slot-N. A slot is just a display position a sample occupies; dragging cards rearranges which sample sits where.

M9 overlap (awareness note — unchanged intent). M9 "slots" (capture-to-slot-N / insert-slot-N, MIDI-bindable, MPC-style) remains explicitly deferred (Daniel, 2026-07-26). The interchangeable substrate L7 builds still eases a future M9 revival — it lays the addressable-position groundwork M9 would sit on — but L7 adds no slot-numbered capture/insert actions and no MIDI bindings. The "plain vs. M9-shaped" sub-fork is closed: plain gap-preserving substrate, per F2 above.

2. Decorative metadata overlay (bars.beats · s.ms)

Each card overlays, on the waveform, two decorative read-outs of capture length:

• bottom-LEFT: length in bars.beats.subdivisions (musical).
• bottom-RIGHT: length in seconds.milliseconds (wall-clock).

Contract: decorative and non-interactive — no hit-test, no hover, no selection role. Drawn via the L1 kit text() in the micro / value-mono type class, in text/dim (or a subtle shadowed variant for legibility over the peaks), subordinate to the waveform. Respects the speed constraint (no animation). Pure formatting helpers (below) are unit-tested; only the kit draw is shell.

Bars.beats source (F1 — SETTLED 2026-07-27: capture-time stamp). bars.beats.subdivisions requires a tempo + time-signature reference. Sample already carries captureTempo (BPM at capture) and lengthBeats, but no time-signature. Decision: capture-time stamp — add captureTimeSigNum / captureTimeSigDenom to Sample + its JSON round-trip, stamped on the capture path (read the project meter at capture via TimeMap_GetTimeSigAtTime — confirmed at build against the SDK). bars.beats.subdivisions renders from the stamped tempo + meter, stable under later project tempo/meter changes — a bank sample outlives the project state it was captured under, matching the existing captureTempo stamp philosophy. This is a capture-path write beyond draw work (its own checkbox in the PLAN, now settled). Old samples with no stamp fall back gracefully (blank musical read-out, or a documented assumed 4/4).

• (rejected) Live project meter at draw time: the label would drift under the card as the project tempo/meter changes, and would be wrong for any sample captured under a different meter than the project's current one.

Formatting helpers (pure, tested). bars.beats.subdivisions from lengthSeconds + captureTempo + capture-time signature; seconds.milliseconds from lengthSeconds. Deterministic; graceful on edge cases (zero length → both read empty/0; missing tempo → blank musical read-out, keep the s.ms read-out).

3. Selection styling — tertiary border replaces inversion

A selected card drops the inverted accent-fill treatment and instead draws the normal cell (Rest or Hover surface) + an accent/tertiary border (pastel purple #C2AAE8); the waveform draws in its normal accent color (the inverted bg/base wave is removed). The four grid-card interaction states stay visually distinct and coherent:

• Selected: normal cell surface + accent/tertiary (purple) 1px border. No fill change, no wave inversion.
• Focus (caret): the existing distinct inner ring — keep it separate from the selection border so a focused and selected card reads both (e.g. purple outer border + a text/primary inner focus ring). Settle the exact inner treatment at build so focus is legible on top of the selection border.
• Hover: the kit Hover surface (unchanged) — a fill-state change, orthogonal to the purple border, so a hovered selected card still reads as selected.
• Drag-target slot: the reorder drop-target highlight (a distinct accent — recommend accent/hot outline on the target slot) must not be confusable with the purple selection border; spec the exact treatment at build.

Drag disambiguation (F3 — SETTLED 2026-07-27)

In-grid reorder must coexist with the existing internal bank-move/copy drag and OS drag-out. Precedence (one clean rule, evaluated live during the drag):

1. Pointer leaves the client rect → OS drag-out (unchanged; drag_out::decideGesture wins first — the existing invariant #4 boundary).
2. Else drop lands on a tab / the OTHER region's bank → move/copy (unchanged; Ctrl = copy).
3. Else drop lands within the SAME bank's own grid → reorder-to-slot (new).
   • Onto an empty slot → place there.
   • Onto an occupied slot, no modifier → insert-before-and-shift-tail (the default, above).
   • Onto an occupied slot, Alt held → REPLACE the occupant (below).

So: leave-client wins → else other-bank wins → else same-bank-grid = reorder. The precedence is encoded in a pure decision helper (mirror drag_out::decideGesture); the shell reads the live pointer + focused region + client rect + modifier state (Alt) and calls it. This keeps the reorder gesture from ever stealing an intended bank-move or OS-drag, and keeps a same-bank in-grid drag from being mis-read as a no-op (today a same-bank drop is a no-op; L7 gives it reorder meaning).

Alt+drop-onto-occupied = REPLACE (SETTLED 2026-07-27). Holding Alt at drop over an occupied slot replaces the occupant instead of inserting-and-shifting. Replace semantics, precisely:

• The replaced sample is removed from THAT bank's index only — same semantics as the existing remove-from-bank verb (BankBook::removeSample index-only). The file stays on disk; the owned-manifest and Phase R prune govern its bytes. If the replaced sample's last reference disappears, that is exactly the existing cross-bank-reference story (hashReferencedElsewhere reports whether the hash is still referenced elsewhere; prune later reclaims a now-orphaned owned file). Replace introduces no new deletion authority — it never touches the disk.
• The dragged sample then takes the vacated slot (the slot's position is preserved; only its occupant changes).
• Pool case (un-evacuable pool — SETTLED rule). The pool's privileges (un-deletable, un-renamable, un-evacuable, never zero banks) are enforced in bank_book. Alt+Replace is allowed in the pool only when it does not violate a pool privilege. Concretely: replacing a pool entry is an index-only removal of that pool sample; it is permitted as long as it does not empty the pool below the pool's floor and does not remove the last reference in a way the pool's rules forbid. The consistent rule the model enforces: the replace's index-removal step is the same operation as remove-from-bank, and it must pass the same pool-privilege guard that remove already applies — if remove-from-pool would be rejected for that sample, Alt+Replace over it is rejected too (the drop falls back to a no-op; the pool-privilege guard is reused as-is — confirmed at build). No special pool-only replace path; one rule, guarded by the existing pool invariants.

Drop-result cursor cues (SETTLED 2026-07-27 — REAPER-idiomatic special cursors). During a drag the cursor must indicate what the drop will do, following REAPER's idiomatic use of distinct action cursors. The cue set:

• Reorder-to-slot (within the same bank's grid) — a move/reorder cursor.
• Move/copy to another bank or tab — the move (or copy, when Ctrl is held) cursor, matching the existing internal-drag semantics.
• OS drag-out (pointer left the client rect) — the OS copy/drag cursor (owned by the OS drag loop once handed off).
• Replace — a distinct replace cursor, shown only when Alt is actually held over an occupied slot (i.e. only when precedence resolves to the Alt+replace case). It must not appear over an empty slot or when Alt is not held.

Shell mechanism: the shell sets the cursor via Win32/SWELL SetCursor. SWELL stock cursors were chosen at build (Reorder→IDC_SIZEALL, Move→IDC_HAND, Copy→IDC_UPARROW, Replace→IDC_SIZEWE; no custom cursor load/synthesis required). The decision of which cue applies stays in the pure gesture/disambiguation helper — the same helper that resolves precedence returns the resolved gesture (reorder / move / copy / os-drag-out / replace), and the shell maps that pure result to a cursor. No cue logic in the shell; the shell only owns the SetCursor call and the cursor resources.

Pure/shell discipline (L7)

Model: the position carrier + gap semantics + reorder mutator + Alt-replace mutator (the latter reusing the existing index-only remove-from-bank semantics + pool-privilege guard) + JSON round-trip/migration are pure (in bank_book, CTest-covered to the bar of its existing round-trip). Layout: the sparse-aware slot↔rect math + point→slot hit-test + the drag-disambiguation decision (including the resolved-gesture result that drives the cursor cue, and the Alt-over-occupied → replace resolution) are pure (extend bank_grid; mirror mode_switch / drag_out::decideGesture). Formatting: the bars.beats and s.ms formatters are pure. Shell (DAW-bound): the reorder/replace drag wiring + drop-target highlight, the cursor SetCursor call mapping the pure resolved-gesture to a cursor resource, the capture-time-signature stamp (F1, settled) read on the capture path, and the kit overlay/selection-border draw. The L1 kit draws; no palette/font decision re-opened.

The L3 gate + Phase S coordination contract

L3 (VST editor + embed-strip restyle) is GATED on Phase S landing on dev. The VST editor (IPlugView LICE surface), the S6 embed strip, and the keyboard strip live in Phase S, which is not on dev yet (it exists on the phase-s worktree). L3 cannot be built on dev until Phase S's editor/embed surfaces (≈ S1 / S6 / S10) merge to dev — the Phase L team must not chase these files on dev; they are not there. Until then L3 is a planned, blocked point; L1, L2, L4, L5, L6, and L7 have all landed — see COMPLETED.md. L4, L5, L6, and L7 are independent of the L3 gate.

Coordination contract (load-bearing): Phase S's S10–S13 build their interaction UX with the current drawing and adopt the L1 kit when it is available — they are NOT gated on Phase L. Whichever lands first (the L1 kit or the S10–S13 UX), the kit is the one source of drawing: if S10–S13 reach dev before L1, they draw in the current language and L3 restyles them; if L1 lands first, they are born in the kit. Either way there is one kit and one look; L3 completes the VST/embed adoption and applies the settled-and-revised B + three-accent pastel treatment (with C's pastel spectral keyboard strip as the signature surface), routing text through the kit's cached-font text() (§3.1 — the kit's current face; no font change). The VST3 class UID is unchanged — a visual refresh is not a compat event.

Precision / invariant implications (what Phase L does NOT change)

• The pure/shell split holds. All layout/hit-test stays in pure CTest-covered geometry modules; the kit's draw half is shell, its geometry half is pure — even where a WDL piece is reused (DS-1). No hit-test math moves into untestable code.
• RT discipline untouched. The kit is draw-thread only; nothing here touches process or any off-thread reload handoff (a Phase S concern surfaced at L3).
• Read-only-over-bank untouched. This is look-and-feel; no data-ownership change.
• The capture/placement load-bearing principle is untouched. Phase L draws; it does not capture, place, or mutate the bank.
• VST3 class UID / component-state contract unchanged (Phase S concern; noted for L3).
• Windows-only (D5). Font/GDI/HFONT choices assume Windows; no cross-platform font fallback concern.

LICE / WDL API surface (verify all signatures)

• LICE design-kit surfaces (L1). LICE_GradRect, LICE_RoundRect, AA LICE_Line/LICE_FLine/LICE_ThickFLine/LICE_Circle/LICE_FillCircle/ LICE_DrawCBezier, LICE_FillTriangle/FillTrapezoid/FillConvexPolygon, and the LICE_CachedFont/LICE_IFont font engine (SetFromHFont, AA DrawText, shadow/outline/ glow FX flags). Verified present in vendor/WDL/WDL/lice/lice.h + lice_text.h; confirm exact signatures + the LICE_CachedFont↔HFONT lifecycle at build.
• WDL/vwnd reuse (DS-1, build-time assessment). virtwnd-slider.cpp / vwnd_slider_drawknobstack (slider/knob drawing reference), virtwnd-listbox.cpp (candidate scroll listbox), virtwnd-controls.h (WDL_STYLE_* gradient hooks), virtwnd-skin.h (image-skin helpers) — all in vendor/WDL/WDL/wingui/. Reuse where a piece beats re-deriving; keep hit-test geometry pure regardless.
• Panel drawing/hit-test (L2). Reuses the bank_panel LICE surface + the existing pure bank_grid/tab_strip/mode_switch hit-test modules; the new action-bar layout is a new pure module. WM_MOUSEMOVE/TrackMouseEvent (WM_MOUSELEAVE) for hover on the panel's existing timer-driven wndProc. Verify against the SWELL headers as bank_panel already does.

Non-goals / guardrails

• No external UI framework. iPlug2 / JUCE / VSTGUI are rejected (DS-1). LICE + reused WDL pieces are the toolkit; reject any path that pulls in a new framework.
• No hit-test geometry in untestable shell code. Even when reusing a WDL piece, layout/ hit-test math stays in pure CTest-covered modules (DS-1 caution). Reject a control whose adoption would move geometry into the shell without a pure test seam.
• L2 does not restructure the panel bones. The vertical-split / grid / tab structure is sound and stays; L2 designs the layout of the button inventory around it (DS-3). A ground-up structural rework of landed Phase-B panel structure is out of scope.
• L3 does not build on dev until Phase S lands there. The gate is explicit; do not chase Phase S files on dev.
• A visual refresh is not a compat event. VST3 class UID, command-id strings, ext-state namespaces, and component-state contracts are unchanged by Phase L.
• Verify LICE/WDL/SWELL surfaces against vendor/WDL before use.

---

Structural reorganization — reorg spec (Phase Q — Quality)

New pillar, own lettered phase, and the LAST structural pillar. Phase Q is a pure structural refactor of src/ — more encapsulation, granular namespaces, core//shell// app/ subdirectories — against a stated quality bar ("mtytel Vital is my code reference for quality"), to bring the codebase "into the realm of something I can stand to look at." It ships no feature and changes no behavior: the test suite passing unchanged is the proof of correctness. Namespaced Q (Quality) — M/D/B/R/V/S/L taken; Q names the end (the quality bar), the reorg being the means. Product framing, the Vital-grounded target shape, the grep-verified SOLID audit (the evidence base), and the fork record (Q-1..Q-6): docs/product/code-organization.md. When a point lands, doc-keeper moves it to COMPLETED.md.

What it is

A directory + namespace + file-split reorganization that makes ReaSampler's already-real, CMake-enforced architecture legible in the code's shape. The pure/shell split exists (30 pure static libs, each with its own test executable, none linking a REAPER SDK) but is invisible: all 45 files sit in one flat src/, all 37 headers in one flat reasampler namespace, four modules have grown into god-modules, and the JSON parser is copy-pasted across four models. Phase Q gives the existing subsystem grouping — model / view / capture / audio / ui / reclaim / version — a structural home (directories + namespaces), splits the four god-modules along validated seams, and extracts the duplicated JSON into one pure module. Nothing about the architecture changes; it becomes visible. This is why the phase can be zero-runtime-cost and CTest-green throughout: the seams already exist in the link graph; Phase Q draws them where a reader sees them.

The quality bar — Vital (read from its actual src/ tree)

Vital (github.com/mtytel/vital) groups its synth by subsystem — common/ synthesis/ interface/ plugin/ — with synthesis/ further subdivided by function (synth_engine/ modulators/ filters/ effects/ producers/ framework/ lookups/ utilities/). The grouping principle is layered functional architecture: the directory tree is the architecture diagram; you read the subsystem map off the folders. ReaSampler adopts the pattern (directory = architecture), adapted to its own most load-bearing invariant — the pure/shell split — as the top level (see below). Vital is GPLv3; the borrowed artifact is the structural pattern, not code.

Settled decisions (Q-1 settled; Q-2..Q-6 recommended — see docs/product/code-organization.md §6)

• Q-1 — namespace letter. SETTLED: Q (Quality). Point-id family Q1..Qn, wave prefixes Q-W1..Q-W6. O (Organization) was set aside: the glyph reads ambiguously against zero in point ids, and "Organization" undersells a phase measured against a quality bar.
• Q-2 — JSON extraction in scope + first. REC: yes. The 4× duplicated Parser is the largest DRY+SRP violation and is entirely off the hot paths — the ideal safe, high-leverage opener (Q-W1).
• Q-3 — directory shape. REC: core//shell//app/ top-split, subsystem dirs beneath. Top level by the pure/shell discipline (so the invariant is structural, not merely conventional), subsystem grouping one level down. Preferred over pure-Vital subsystem-first because ReaSampler has a pure/host split Vital lacks and that split is the invariant most worth protecting structurally.
• Q-4 — sub-namespace to match sub-directory. REC: both. reasampler::model/view/capture/ audio/ui/reclaim/version/json. Directory and namespace agree; a symbol's home is unambiguous from either.
• Q-5 — god-module split granularity. REC: to the audit's named seams, no finer. Well-factored, not atomized.
• Q-6 — OCP registration-table. REC: in scope, last (most droppable if narrowing).
• Q-7 — naming rides the relocation waves, no dedicated naming wave. REC: yes (forced once Q-3/Q-4 settle — a rename is near-free during relocation, near-pure-churn standalone).
• Q-8 — class/module renames beyond the free namespace fix. REC: fix the two that actively mislead — BankIndex→BankModel (the bank_model.h/BankIndex file↔class word-mismatch) and the unified JSON parser → json::Reader/json::Writer (or json::Parser) — leave the merely-quirky (Book/Bank/Index, Sample/AudioSample, MinMax, KitBox). Daniel's to call.
• Q-9 — align the capture_realtime (shell) / realtime_record (pure) word-order inversion. REC: yes, during Q-W3 (a free rider — W3 already hoists the realtime lifecycle).

The directory + namespace map (Q-3 / Q-4)

Top-level by the pure/shell discipline; subsystem dirs beneath core/; namespaces mirror directories.

core/ (pure, no REAPER types, unit-tested — reasampler::<subsystem>):

• core/model/ (::model) — bank_model, bank_book, owned_manifest, provenance
• core/view/ (::view) — view_mode_model, view_tree, lane_keys, mode_switch
• core/capture/ (::capture) — render_settings, batch_capture, tail_control, capture_paths, wav_trim, insert_plan
• core/audio/ (::audio) — peaks
• core/ui/ (::ui) — theme, component_geometry, bank_grid, tab_strip, action_buttons, prune_button
• core/reclaim/ (::reclaim) — prune_reconcile
• core/version/ (::version) — app_version
• core/json/ (::json) — NEW — extracted parser/serializer (replaces the 4 duplicate Parsers)

shell/ (REAPER-facing — subdir by subsystem, namespace as house style prefers):

• shell/capture/ — capture, capture_realtime, provenance_shell, track_guid, item_read; post-Q-W3 capture_orchestrator, scope_resolve, realtime_lifecycle
• shell/panel/ — draw_kit; post-Q-W2 panel_render, panel_thumbnails, panel_audition, panel_input, panel_bank_ops, panel_window (from bank_panel)
• shell/view/ — view
• shell/persist/ — post-Q-W5 session, ext_state_io, prune_fs (from persist)
• shell/actions/ — drag_out_win; post-Q-W4 design_view_actions, bank_actions, prune_action (from actions)

app/: main.cpp (post-Q-W3: API-pointer ownership + ReaperPluginEntry + dispatch only).

Collision sweep (before W1): Sample (::model) vs AudioSample (::audio alias) vs the unified Parser (::json) must not collide once flattened into granular namespaces; resolve by subsystem home.

The god-module split seams (Q-5)

Split each to the audit-validated seams, no finer:

• bank_panel.cpp (2424 LOC → shell/panel/, Q-W2): panel_render (draw/paint) / panel_thumbnails (compute+cache) / panel_audition (preview engine — hot path, direct call-through) / panel_input (mouse/key/wheel + new-content detection) / panel_bank_ops (bank CRUD — the single owner W4 dedupes against) / panel_window (lifecycle + OS drag-out/drop-target). Split the fat bank_panel.h per seam (I).
• main.cpp (1762 LOC → hoist to shell/capture/, Q-W3): capture_orchestrator (RunCapture/captureAndIndexOne/renderOffline/batch/recapture/realtime Run*) / scope_resolve (range/razor/track resolution + provenance assembly inputs) / realtime_lifecycle (state machine + globals + selection guards). FxBypassGuard moves out but stays stack RAII (precision-critical). main.cpp → app/, reduced to pointers + entry
  • dispatch.
• actions.cpp (981 LOC → shell/actions/, Q-W4): design_view_actions / bank_actions / prune_action (doBankPruneFolder — the single file-deletion action). Dedupe promptText/mintBankId + bank verbs against panel_bank_ops.
• persist.cpp (766 LOC → shell/persist/, Q-W5): session (lifecycle/poll + BeginLoadProjectState reload hook) / ext_state_io (serialization bridge + GUID minting + folder relocation) / prune_fs (prune scan + deleteOrphanFile via SHFileOperationW — the isolated single file-deletion authority).

The JSON extraction (Q-2 / Q-W1)

Extract one pure core/json (::json): parser (parseString/parseInt/parseKey/ skipValue + escape) + serialize/emit helpers. Rewire bank_model, bank_book, view_mode_model, owned_manifest onto it and delete their four hand-rolled Parsers. Round-trip output must be byte-identical to before — this is a structural dedupe, not a format change. Off all hot paths (serialization runs at save/load, never per frame) — safe to abstract freely.

The OCP registration-table (Q-6 / Q-W6)

Replace the ~350-line hand-written non-table action registration blocks (isolated in app/main.cpp after Q-W3) with a data-driven registration table, so adding an action edits one place, not four parallel Register("command_id"/"gaccel"/"hookcommand") sites. Unload mirror-unregisters from the same table. Command-id + display strings stay byte-identical (FOREVER-STABLE, per-channel — the Phase V V4 contract). Also split residual fat headers (capture.h/persist.h) alongside their TU splits (I).

The naming dimension (Q-7 / Q-8 / Q-9 — grep-verified audit in docs/product/code-organization.md §2b)

Alongside giving symbols a directory + namespace home (Q-3/Q-4), Phase Q gives inconsistently/poorly-named symbols a consistent name — same Vital bar, orthogonal to the SOLID focus. The audit (2026-07-27) is grep-verified; the load-bearing findings:

• Already consistent — preserve verbatim: the geometry-mirror verb vocabulary (compute<Thing>Rects/compute<Thing> + hitTest<Thing>, verified across bank_grid / mode_switch / action_buttons / action_bar / tab_strip / prune_button / overflow_menu / footer_bar / card_drag / component_geometry) and the uniform <module>_tests CTest suffix.
• Collisions — resolved by the Q-4 sub-namespaces for free: four class Parser (bank_model.cpp / bank_book.cpp / owned_manifest.cpp / view_mode_model.cpp) collapse to one json::Parser in Q-W1; the shared pure-UI rect types FooterRect / ButtonRect (defined in prune_button.h, reused by footer_bar.h under an explicit hand-collision "NAME NOTE") get one ui:: owner; Sample (model::) vs AudioSample (audio::) de-collide by home.
• Genuine renames (Q-8/Q-9 — Daniel's call): BankIndex→BankModel (the bank_model.h file↔class word-mismatch — the worst legibility wart, rec: rename the class so the model family reads BankModel/BankBook/ViewModeModel); the unified JSON parser named json::Reader/ json::Writer at W1 mint; align capture_realtime(shell)/realtime_record(pure) to the house shell↔core convention (drag_out↔drag_out_win is the model) during Q-W3.
• Sequencing (Q-7): renames ride the wave that already relocates/splits the file — no dedicated naming wave. Rule: no rename lands on a file the wave isn't otherwise touching. W1 carries the collision + model-class renames; W2 the panel_* names; W3 the realtime word-order fix. Zero-behavior-change like the rest of Phase Q; verified by the module's own test executable.
• Out of scope (never renamed): the FOREVER-STABLE contract strings are not C++ symbols — command_id strings, action display names, ext-state namespace ("reasampler"/"reasampler_beta") and keys, the reasampler: lane prefix, the VST3 class UID. Renaming a C++ class is orthogonal to these literals (audit §2b.5).

Performance guardrails (HARD CONSTRAINT — Daniel's non-negotiable)

The reorg must cost zero runtime. The two hot paths must keep their exact call/inline shape; the following are acceptance criteria on every point:

• peaks envelope compute (computeEnvelope / lastFrameAboveThreshold over full PCM): NO virtual dispatch, NO peaks interface, NO added header→TU indirection. computeEnvelope stays a free function on const std::vector<float>& so it inlines as today. Relocate + namespace only; never wrap in an abstraction.
• Audition / preview: panel_audition may be its own TU (Q-W2), but the call stays a direct call-through, not virtual.
• Realtime-capture tick: keep the idle fast-path a single pointer test; realtime_lifecycle (Q-W3) must not change the tick's branch shape.
• JSON extraction is off all hot paths — safe to abstract freely (why Q-W1 is the safe opener).
• FxBypassGuard runs per-capture, not per-frame — keep it stack RAII when it moves out of main.cpp (Q-W3); never heap-allocate or virtualize it.

Net: every recommended split falls on a cold path or preserves call/inline shape on the two hot ones. A split that would add a hot-path indirection is out of scope — rework it or drop it.

The GATE (load-bearing — Phase Q is last)

Phase Q is gated on the tree being otherwise quiescent. Daniel's plain readiness target: "when Phase S and L3 are finished." As of 2026-07-27 the outstanding work is precisely (1) Phase S merged to dev (the large second-artifact branch, currently on the phase-s worktree) and (2) Phase L L3 merged to dev (the VST restyle, itself gated on Phase S). L1/L2/L4/L5/L6/L7 have already landed (see COMPLETED.md) — the earlier "L2 + L3" wording was stale and is corrected to L3 only. D2 is functionally complete (D2-W1..W3-B landed; its lone open item, a per-track lane-split panel indicator, is explicitly deferred, not a blocking residual). M9 (slots) is explicitly deferred (Daniel, 2026-07-26), not scheduled work. D2 and M9 are named here only so that reactivating either re-arms the quiescence condition; neither blocks the gate today. Why: Phase Q touches nearly every file in src/ (relocate, re-namespace, split the four largest TUs, plus the §2b renames); every large in-flight branch (Phase S on its worktree, L3 once it lands) is diffed against the current flat layout, so landing a rename-and-relocate-everything reorg mid-flight forces every open branch through the worst conflict class — a combinatorial re-resolution, not a linear one. Phase Q is last precisely because it reshapes the ground every other pillar stands on: landing it early taxes every subsequent phase; landing it last taxes nothing. Re-confirm quiescence against dev before Q-W1.

M9 disposition (Daniel-decision note). "Deferred indefinitely" ≠ "abandoned." Immaterial to the gate (both clear it); matters only if M9 is reactivated — before Phase Q it lands cheaply on the flat layout, after it is authored against the reorganized tree. Surfaced, not silently resolved (full note: docs/product/code-organization.md §4).

Wave sequencing (each independently landable, CTest-green at every step)

Big-bang is rejected — the CMake per-module static-lib + per-module test-executable seams make a file move + namespace change mechanically verifiable (ctest --test-dir build green or not, at every commit), a property only an incremental reorg uses. Risk-ordered:

• Q-W1 — safe opener: core/json extract (delete 4 Parsers) + impose the directory/ namespace layout on the 30 clean pure libs + clean shells (pure relocation, no logic change). All later waves assume this layout. Carries the naming collision fixes + the model-class renames (Q-8), which are free during this relocation.
• Q-W2..Q-W5 — the four god-module splits, one per wave, risk-ordered (bank_panel → main.cpp → actions.cpp → persist.cpp). Q-W4 depends on Q-W2 (panel_bank_ops dedupe target); Q-W5 best after Q-W4 (prune_action → prune_fs routing); otherwise parallel-safe. Q-W2 carries the panel_* names; Q-W3 carries the capture_realtime/realtime_record word-order fix (Q-9).
• Q-W6 — OCP registration-table + residual fat-header (I) splits. Depends on Q-W3 (registration code isolated first). Sequenced last; most droppable if narrowing.
• Naming (Q-7): no dedicated wave — every rename rides the wave already relocating/splitting its file; a rename that would touch an otherwise-untouched file is deferred.

Precision / invariant implications (what Phase Q does NOT change)

• The pure/shell split is strengthened, never dissolved. Top-level core/ vs shell/ makes it structural; no file crosses the boundary; no core/ file gains a REAPER type; the CMake per-module test executables stay green.
• Every precision invariant holds. Null test, bit-identical repeats, non-destructive capture, exact bounds — none is code Phase Q rewrites. FxBypassGuard moves but stays stack RAII with identical behavior.
• Capture ≠ placement. No hoisted Run* / capture_orchestrator path may gain an InsertMedia call during the split.
• Single file-deletion authority is concentrated, never spread. prune_fs (Q-W5) is the one module that deletes bytes; the reorg improves this invariant.
• Relative-paths-only in the persisted index — untouched.
• On-the-wire/on-disk contract strings unchanged. Command-id strings, action display names, ext-state namespaces, VST3 class UID are byte-identical (per-channel, per Phase V V4). Re-namespacing C++ symbols is orthogonal to these.
• No behavior change. Phase Q is a pure structural refactor; the test suite passing unchanged is the proof of correctness. A point that changes observable behavior has exceeded its charter.

Non-goals / guardrails

• No feature, no behavior change. If a point ships anything user-visible, it is out of scope.
• No hot-path indirection. No virtual dispatch or header→TU indirection on peaks/audition/ realtime-tick — ever (a hard acceptance bar, not advice).
• No design-level (D-letter SOLID) rework. main/bank_panel depending on concrete capture backends is a low-priority Dependency-Inversion concern — out of scope (a design change, not a reorg). Phase Q reorganizes; it does not re-architect interfaces.
• No peaks data-ownership change. peaks forcing a whole-file std::vector<float> copy on the thumbnail path is noted but not touched — reworking it risks the hot path.
• No big-bang commit. Every wave is independently landable and CTest-green; reject a change set that cannot be verified at each step.
• Do not begin before the GATE. Re-confirm the tree is quiescent (Phase S + L + D2 + M9 merged/closed) before any Q point.
• Verify the CMake src/ path updates and the SWELL/LICE surfaces still resolve after relocation, as the existing build already requires.