docs: spec Phase 21 — windowed streaming buffer for bounded client memory

2026-06-23 00:14:44 -04:00
parent 2c1571057a
commit a84a99c309
2 changed files with 428 additions and 0 deletions
@@ -443,6 +443,85 @@ not the same work; this phase does not satisfy or depend on that one.
 ---
 ## Phase 21 — Windowed Streaming Buffer (bounded client memory for long streams)
 Bound the **client memory** a playing track consumes to a small, configurable forward window —
 **independent of total stream length** — so a 1 GB+ DJ MIX (Phase 9 `Mix` medium: a single long track)
 plays without the whole decoded PCM accumulating in the browser. **Public listener site only**
 (`DeepDrftPublic.Client` player stack + `DeepDrftPublic` TypeScript audio interop); no CMS, no API
 endpoint, no schema change.
 The network path already streams in adaptive 16–64 KB chunks. The accumulation is on the **decode
 side**: `PlaybackScheduler` holds an `AudioBuffer[]` it **never evicts** ("Supports pause/resume/seek by
 retaining all buffers" — its own doc comment). Decoded PCM is larger than the source (Web Audio is
 32-bit float per sample/channel — a 16-bit stereo WAV roughly doubles once decoded), so a 1 GB WAV
 becomes ~2 GB of retained float data. That is the OOM. The fix: hold only a sliding forward window plus a
 small back-retain, discard already-played buffers, and refill on demand.
 **Architectural spine — a sliding window keyed on playback position, built as a generalization of the
 landed seek-beyond-buffer path.** The Phase 4 HTTP `Range: bytes=X-` → 206 primitive already does every
 plumbing primitive the window needs (discard-buffers-keep-offset via `clearForSeek`/`setPlaybackOffset`;
 fetch-from-offset via `TrackMediaClient`; decode-header-less-body via
 `StreamDecoder.reinitializeForRangeContinuation`; time→byte via `IFormatDecoder.calculateByteOffset`),
 just triggered manually and one-shot. The only genuinely new mechanisms are **partial eviction** on the
 scheduler and **back-pressure** on the forward read loop (stop calling `ReadAsync` above a high-water
 mark, resume below low-water). Recommended **Direction A** (sliding window on the existing single forward
 stream); **Direction B** (discrete Range-fetched segments — the HLS/DASH/MSE-eviction analogue) held as
 the documented fallback; **Direction C** (adopt MSE and let the browser manage the buffer) flagged as the
 real long-term answer but out of scope — it is a playback-substrate rewrite entangled with non-WAV
 formats (Phase 1.2), surfaced as OQ5.
 **Invariants that must hold (the §3.5 seam contract).** Reuse the Range path, don't fork it; playback-
 start latency at parity; the `IFormatDecoder` abstraction untouched (windowing is format-agnostic, so
 wiring MP3/FLAC later inherits it free); read-only playback (no new control); the single-instance JS
 decoder stays single-writer (every refill routes through the existing cancellation/drain discipline). The
 **Mix visualizer is provably unaffected** — it renders from the preprocessed per-track high-res datum
 (Phase 10/12), never from live decoded PCM, so evicting played buffers cannot starve it. The 1 GB mix is
 both the canonical case *and* the proof the eviction is safe.
 **Interaction with deferred Phase 1 features (same seam):** windowing should land **before** preload
 (1.3) — it makes preload of long tracks memory-safe by construction (a staged next-track decoder inherits
 the bounded scheduler); it makes crossfade (1.4) between two long mixes affordable (the overlap doubles
 the *window*, not the track); it adds a minor "don't evict the final window before the gapless boundary"
 care point for 1.5. It **enlarges the error surface** (1.6): windowed refill issues mid-stream fetches
 the listener didn't initiate, one of which can fail deep into a 1 GB mix — so the *cheap* half of 1.6
 (clean refill-failure handling, no wedged player) is folded into this phase's acceptance criteria, not
 left fully to 1.6.
 Full design, the three directions with SOLID/road-not-taken rationale, use cases, acceptance criteria,
 the open-question set, and the wave decomposition: `product-notes/phase-21-windowed-streaming-buffer.md`.
 Sequenced as four waves. `21.1 → 21.2 → 21.3`, with `21.4` validating the whole. **21.1 is the cold-start
 prerequisite and the load-bearing change** — independent of the open questions (window *sizes* are
 parameters fed in later).
 - **21.1 — Partial eviction in `PlaybackScheduler` (cold-start; load-bearing).** Drop already-played
  buffers while keeping the position/index/time-anchor bookkeeping exact against a buffer array that no
  longer begins at absolute time 0 (today `getCurrentPosition`/`playFromPosition`/the schedule loop all
  assume `buffers[0]` is the track start). The hardest correctness work in the phase. No refill yet.
  **Independent of the open questions — can begin immediately.**
 - **21.2 — Back-pressure on the forward read loop.** Stop `ReadAsync` above the high-water mark, resume
  below low-water; together with 21.1 this bounds *both* the played and unplayed regions (the AC1
  guarantee). Routes resume/pause through the existing single-loop cancellation discipline. **Depends on
  21.1.**
 - **21.3 — Seek-back-past-window refill.** When a backward seek lands earlier than the retained tail,
  refetch via the existing seek-beyond-buffer Range path pointed at the earlier offset; plus the minimal
  clean refill-failure handling (the 1.6 adjacency). Mostly reuse of the landed seek path. **Depends on
  21.1 + 21.2.**
 - **21.4 — Validation against the 1 GB target (acceptance).** Memory profiling (bounded under 1 GB is the
  headline), latency parity, edge-to-edge playback, the seek matrix, induced refill failure, visualizer-
  running, rapid-seek concurrency. Largely measurement; breaks are tuning fixes in 21.1's anchor math or
  21.2's water-marks. **Depends on 21.1–21.3.**
 **Dependency shape:** `21.1 → 21.2 → 21.3 → 21.4`; 21.1 is the only cold-start wave. **Open questions for
 Daniel (spec §6):** window-size policy axis (time-based window + memory guard — recommended); seek-back-
 past-window re-buffer acceptable (recommend yes, symmetric to forward); a hard total in-flight memory cap
 as a guard rail (recommend yes); window everything vs. only long tracks (recommend everything — one path,
 short tracks never hit a refill); and whether MSE is the real destination (steer informing scope, not a
 blocker). None block 21.1.
 ---
 ---
 ## Working with this file
@@ -0,0 +1,349 @@
 # Phase 21 — Windowed Streaming Buffer (bounded client memory for long streams)
 Product spec. Status: **design / framing — implementation-ready pending Daniel's open-question calls.**
 Author: product-designer. Date: 2026-06-23. **No code has been written by this doc.**
 Surface: **public listener site only** (`DeepDrftPublic.Client` player stack + `DeepDrftPublic`
 TypeScript audio interop). No CMS (`DeepDrftManager`) change. No data-model or schema change. The one
 server touch is **reuse, not new surface**: the existing `DeepDrftAPI` HTTP `Range: bytes=X-`
 partial-content primitive (Phase 4, landed) is the load-bearing dependency; this phase adds no new API
 endpoint.
 ---
 ## 1. Goal
 Bound the **client memory** a playing track consumes to a small, configurable forward window —
 **independent of total stream length** — so a 1 GB+ DJ MIX (Phase 9 `Mix` medium: a single long track)
 plays without the whole decoded PCM accumulating in the browser.
 **The defect, stated precisely.** The network path already streams in adaptive 16–64 KB chunks
 (`StreamingAudioPlayerService.StreamAudioWithEarlyPlayback`) — that part is fine. The accumulation is on
 the **decode side**: `PlaybackScheduler` holds `private buffers: AudioBuffer[]` and **never evicts**
 ("Supports pause/resume/seek by **retaining all buffers**" — its own doc comment). Every 64 KB segment
 the `StreamDecoder` decodes is pushed via `addBuffer()` and kept for the life of the track. Decoded PCM
 is **larger than the compressed-or-raw source** in memory (Web Audio `AudioBuffer` is 32-bit float per
 sample per channel — a 16-bit stereo WAV roughly **doubles** in size once decoded), so a 1 GB WAV becomes
 ~2 GB of retained `AudioBuffer` float data. That is the OOM.
 **One-line framing:** today the player decodes the whole track into memory and keeps it; Phase 21 makes
 it keep only a sliding forward window and discard what has already played, refilling on demand from the
 Range primitive it already uses for seek.
 ---
 ## 2. Constraints / invariants (the contract that must hold)
 These are non-negotiable. The §3.5 streaming seam (root `CLAUDE.md` "Streaming-first audio playback";
 `CONTEXT.md §3.5`) is called *the most architecturally load-bearing part of the playback path* by both
 docs. This phase **modifies that seam** — so the contract it must preserve is spelled out here.
 - **C1 — The seek-beyond-buffer Range path is the substrate, kept intact.** Phase 4 landed HTTP
  `Range: bytes={offset}-` → `206 Partial Content` end to end (client `TrackMediaClient` →
  `DeepDrftPublic` proxy → `DeepDrftAPI`), and `StreamDecoder.reinitializeForRangeContinuation` retains
  the parsed format header on a continuation body (no re-parse). Windowed refill is a **generalization of
  this exact path** (§3.1) — it must not require a second, divergent fetch mechanism.
 - **C2 — Playback start latency unchanged.** Today playback starts as soon as a configurable minimum
  buffer count is queued (header-derived duration, not full-file). The window model must keep first-audio
  latency at parity — bounding memory must not reintroduce a fetch-then-play stall.
 - **C3 — The format-decoder abstraction is untouched.** `IFormatDecoder` (WAV active; MP3/FLAC
  implemented, not yet wired) owns all format-specific byte math. Windowing lives in the
  **format-agnostic** layer (`PlaybackScheduler` eviction + `StreamDecoder`/player refill
  orchestration); it must add **no** format-specific branches. A future wired MP3/FLAC decoder inherits
  windowing for free.
 - **C4 — Read-only playback only.** This is a memory-management change, not a UX change. No new
  user-visible control, no change to seek/transport semantics beyond what the listener already
  experiences. Seek must still feel identical.
 - **C5 — WAV-only is the shipping target; the design must not foreclose MP3/FLAC.** Byte↔time mapping
  for refill is exact and cheap for WAV (CBR: `byteRate` from the header). For VBR formats the mapping is
  approximate (the decoders already carry TOC/SEEKTABLE seek math). The window machinery must express
  refill in terms of the decoder's existing `calculateByteOffset`, so the same code works when those
  formats are wired — **no WAV-special-cased offset math in the window layer.**
 - **C6 — No regression to the single-instance JS decoder concurrency guarantees.** The current code is
  careful that only one streaming loop touches the single JS `StreamDecoder` at a time
  (`DrainActiveStreamingTaskAsync`, the `_streamingCancellation` identity dance). Windowed refill
  introduces *more* mid-stream fetches; it must route through the **same** drain/cancellation discipline,
  not around it.
 - **C7 — The Mix visualizer's data source is independent and must stay that way.** The Phase 10/12
  WebGL2 lava visualizer renders from a **preprocessed high-res waveform datum** fetched per-track
  (`GET api/track/{entryKey}/waveform/high-res`), **not** from live decoded PCM. Confirmed: evicting
  played `AudioBuffer`s cannot starve the visualizer — it never read them. The window model is invisible
  to the visualizer. (This is the canonical 1 GB case *and* the case that proves the eviction is safe.)
 ---
 ## 3. Architectural shape
 ### 3.0 The mental model
 A track's audio is a byte range `[0, fileLength)` on disk. At any moment the listener is at playback
 position `P` (seconds → byte offset via the format decoder). The player should hold decoded
 `AudioBuffer`s only for a bounded window roughly `[P - back, P + ahead]`:
 - **forward fill (`ahead`)** — enough decoded lookahead that playback never starves (covers the existing
  500 ms scheduler lookahead plus network jitter headroom);
 - **back-retain (`back`)** — a small amount of *already-played* audio kept so a short seek-back does not
  trigger a network refetch;
 - **evict** — anything older than `P - back` is dropped (`AudioBuffer` references released → GC reclaims
  the float data);
 - **refill** — when forward decoded lookahead drops below a low-water mark, fetch+decode more from the
  current byte position; when the window's tail is evicted and the listener seeks back past it, refetch
  that region via the Range primitive (the seek-beyond-buffer path, run *backwards*).
 This is a **ring/sliding-window buffer keyed on playback position**, driven by high/low-water marks —
 the standard bounded-producer/bounded-consumer pattern, transplanted onto the decode→schedule seam.
 ### 3.1 Why this is a generalization of seek-beyond-buffer, not a new mechanism
 The seek-beyond-buffer path already does **every primitive** the window needs, just triggered manually
 and one-shot:
 | Window operation              | Existing seek-beyond-buffer machinery it reuses                                    |
 |-------------------------------|-----------------------------------------------------------------------------------|
 | Discard buffers, keep offset  | `PlaybackScheduler.clearForSeek()` + `setPlaybackOffset()` (clears buffers, retains the absolute-time anchor) |
 | Fetch from a byte offset      | `TrackMediaClient.GetTrackMedia(key, byteOffset)` → `Range: bytes=X-` → 206       |
 | Decode a header-less body     | `StreamDecoder.reinitializeForRangeContinuation(remainingByteLength)`              |
 | Map time → byte offset        | `StreamDecoder.calculateByteOffset()` → `IFormatDecoder.calculateByteOffset()`    |
 | Single-loop safety on refetch | `_streamingCancellation` swap + `DrainActiveStreamingTaskAsync()`                  |
 The difference is **eviction does not exist yet** (the scheduler only ever `clear()`s wholesale) and
 **refill is one-shot** (a seek, not a continuous low-water-triggered loop). So the new work is two
 seams: a *partial-evict* on the scheduler, and a *position-driven refill controller* on the player. The
 fetch/decode/offset plumbing is reused verbatim.
 ### 3.2 The three candidate directions
 Per file convention the alternatives are recorded; the recommendation follows.
 **Direction A — Sliding window on the existing single forward stream (recommended).**
 Keep the current model where the C# loop reads one forward HTTP stream and pumps chunks into the JS
 decoder. Add two things: (1) `PlaybackScheduler` gains *partial eviction* — drop buffers whose
 absolute-time end is older than `P - back`, adjusting its index bookkeeping so `getCurrentPosition()`
 and scheduling stay correct against a buffer array that no longer starts at index 0; (2) a
 *back-pressure* signal — when forward decoded lookahead exceeds the high-water mark, the C# loop
 **pauses reading** the HTTP stream (stops calling `ReadAsync`) until playback drains it below low-water,
 then resumes. Memory is bounded by high-water + back-retain. Seek-back beyond the retained window falls
 through to the **existing** seek-beyond-buffer path unchanged.
 *Why recommended:* smallest change to the load-bearing seam; reuses the live forward stream (no extra
 connections in the common case); eviction and back-pressure are the only genuinely new mechanisms, and
 both are local (one to the scheduler, one to the read loop). Back-pressure via "stop reading the socket"
 is exactly how TCP flow control already wants to behave — pausing `ReadAsync` lets the kernel window
 close; we are not fighting the transport.
 **Direction B — Discrete window segments, each its own Range fetch.**
 Treat the file as fixed-size byte segments (e.g. 4 MB). Hold N decoded segments around `P`; fetch the
 next/previous segment via a fresh Range request as the window slides; discard the far segment. No live
 long-lived forward stream — every window is an independent 206.
 *Why not (default):* turns one connection into many short Range requests (more proxy hops through
 `DeepDrftPublic`, more server-side `WavOffsetService`-style header synthesis, more places a fetch can
 fail mid-stream — worsening the §1.6 error surface), and the byte↔time segment math must be exact at
 every boundary. It *is* the cleaner model for true random-access (and the better base if seeking-heavy
 usage dominates), so keep it as the fallback if Direction A's back-pressure proves leaky in practice.
 Borrowed prior art: HLS/DASH segment windows and the MSE `SourceBuffer.remove()` eviction model — this
 is how every production HTML5 adaptive player bounds memory. We are doing the hand-rolled equivalent
 because the stack is a bespoke Web Audio graph, not `<media>` + MSE.
 **Direction C — Adopt MediaSource Extensions (MSE) and let the browser manage the buffer.**
 Stop hand-rolling the decode→schedule graph for long tracks; feed the Range stream into a `SourceBuffer`
 and let the browser evict via its built-in quota + `remove()`. Memory management becomes the platform's
 problem.
 *Why not (now, but flag for Daniel):* MSE does not accept raw WAV/PCM — it wants containerized formats
 (fragmented MP4/WebM, or MP3/AAC elementary streams). The current producer is WAV-only, and the entire
 bespoke visualizer/spectrum graph is wired to the Web Audio `AudioContext`, not a `<media>` element.
 Adopting MSE is a **rewrite of the playback substrate**, not a windowing change — out of scope for this
 phase. But it is the *real* long-term answer and is entangled with Phase 1.2 (non-WAV formats): if
 DeepDrft moves to a compressed delivery format, MSE becomes viable and could retire the hand-rolled
 decoder, the seek-beyond-buffer path, *and* this phase's window machinery in one move. **Surfaced as
 open question OQ5** — not to decide now, but so this phase is built knowing it may be superseded.
 ### 3.3 Recommended direction: A, with B held as the documented fallback
 Direction A is the smallest coherent change that hits the headline (bounded memory under a 1 GB stream)
 while honoring C1–C7. It keeps the live forward stream, reuses the seek-beyond-buffer path for the only
 genuinely random-access case (seek-back past the retained tail), and isolates the two new mechanisms.
 **The final architecture and the exact eviction/back-pressure API are staff-engineer's call at
 implementation** (per file convention); this spec fixes the *shape* and the invariants, not the method
 signatures.
 ### 3.4 SOLID / road-not-taken rationale
 - **SRP, preserved.** Eviction is a `PlaybackScheduler` concern (it already owns buffer storage); refill
  orchestration is a player-service/`StreamDecoder` concern (they already own the fetch loop); byte↔time
  math stays in `IFormatDecoder`. No responsibility crosses a boundary it does not already own.
 - **OCP, via C3/C5.** Windowing added in the format-agnostic layer means wiring MP3/FLAC later changes
  zero window code. The window expresses refill through `calculateByteOffset` — the one seam the
  decoders already implement.
 - **The seam stays single-writer (C6).** Every new refetch routes through the existing
  cancellation/drain discipline, so "only one loop touches the JS decoder" remains true. This is the
  rule most likely to be violated by a naive implementation and is called out as a hard invariant.
 - **Road not taken — eager full decode with a memory cap that just stops decoding.** Tempting (decode
  until you hit a byte budget, then stop) but it breaks playback of long tracks past the cap entirely —
  it bounds memory by *refusing to play the rest*, not by sliding. Rejected: it is a degradation, not a
  feature.
 ---
 ## 4. Use cases
 - **UC1 — Play a 1 GB+ DJ MIX start to finish (the headline).** Memory stays bounded throughout; the
  listener experiences continuous playback identical to a short track.
 - **UC2 — Seek forward within a long track.** Already handled by seek-beyond-buffer; under windowing the
  forward seek clears the window and refills at the target — no behavior change, now with eviction so the
  pre-seek region does not linger.
 - **UC3 — Seek back a few seconds.** Served from the back-retain window with **no** network refetch
  (the reason `back` exists).
 - **UC4 — Seek back far, past the evicted tail.** Falls through to the existing seek-beyond-buffer Range
  fetch, run toward an earlier offset. (Open question OQ2 — see §6.)
 - **UC5 — Pause a long track for a long time.** Memory stays at the bounded window size while paused (no
  continued decode). On resume, forward fill restarts from the low-water trigger.
 - **UC6 — Mix detail page with the lava visualizer running.** Visualizer reads its preprocessed datum
  (C7); windowing is invisible to it. Confirmed non-interaction.
 ---
 ## 5. Interaction with the deferred Phase 1 streaming features
 This phase touches the **same decoder/scheduler seam** as the deferred Phase 1.3/1.4/1.5 items and the
 1.6/1.7 robustness items. The interactions, explicitly:
 - **1.3 Preload / prefetch (deferred; preload half).** *Shares machinery, does not conflict — and should
  be sequenced after.* Preload stages the **next track** into a second decoder instance during the
  current track's tail; windowing bounds the **current track's** forward buffer. They are orthogonal
  axes (next-track vs. current-track-window), but they compound the memory question: a naive preload of a
  second 1 GB mix would reintroduce the OOM this phase fixes. **Recommendation: land windowing first**,
  so that when preload arrives, the staged next-track decoder is *also* windowed by construction (it
  inherits the bounded scheduler). Windowing makes preload *safe for long tracks*; without it, preload of
  mixes is a memory hazard.
 - **1.4 Crossfade (deferred).** Needs two simultaneous `PlaybackScheduler` instances briefly overlapping.
  Both would be windowed instances — the overlap doubles the *window* size momentarily, not the whole
  track. Windowing makes crossfade between two long mixes affordable. No reordering needed; 1.4 still
  gates on 1.3.
 - **1.5 Gapless (deferred).** Sample-accurate hand-off of the next track's first buffer at the current
  track's last buffer. Windowing changes *which* buffers are retained but not the hand-off mechanism;
  the only care point is that the current track's **final** window must not be evicted before the gapless
  boundary is scheduled. A minor invariant for whoever builds 1.5, not a blocker. Note 1.5's existing
  WAV-only caveat stands.
 - **1.6 Track-skip on error (deferred).** *Windowing enlarges the error surface — call this out.* Today
  a fetch failure happens at load (one fetch) or at a user seek (one fetch). Windowed refill issues
  **mid-stream** fetches the listener did not initiate; one of those can fail at byte 700 M of a 1 GB
  mix. So Phase 21 should ship with at least the *cheap* half of 1.6: a mid-stream refill failure must
  **surface a clear error and not wedge the player** (it must not leave playback "running" with a starved
  scheduler — mirror the `playFromPosition` end-of-buffer recovery already in `PlaybackScheduler`). The
  rich half (byte-scan to next valid frame) stays deferred. **Recommendation: fold the minimal refill-
  failure handling into Phase 21's acceptance criteria** (AC6) rather than leaving it entirely to 1.6 —
  it is created by this phase.
 - **1.7 Safari compatibility (deferred).** Windowing adds no new Safari-specific surface beyond what the
  streaming path already has. The one adjacency: more frequent `AudioContext` activity during refill
  should be checked against the older-Safari `webkitAudioContext` quirks when 1.7 is addressed — note it,
  do not block on it.
 ---
 ## 6. Open questions for Daniel (genuine product decisions, not implementation detail)
 These are policy calls with user-visible or resource trade-offs — flagged rather than decided here.
 - **OQ1 — Window size policy.** What bounds the window — a **fixed byte/time budget** (e.g. "hold at
  most ~30 s decoded ahead + ~10 s behind"), or a **configurable memory budget** (e.g. "≤ N MB of
  decoded PCM") that derives the time window from the stream's byte rate? Recommend a **time-based
  forward window + small time-based back-retain** as the primary knob (intuitive, format-portable), with
  a hard **memory ceiling** as a secondary guard. The exact numbers are tunable post-landing; Daniel
  picks the *policy axis*. `[Daniel decision]`
 - **OQ2 — Seek-back past the evicted window.** When the listener seeks back earlier than the retained
  tail, we must refetch (the audio is gone). Acceptable to take the same brief re-buffer the forward
  seek-beyond-buffer takes today? (Recommend yes — it is the symmetric case and listeners already accept
  it forward.) Or should back-retain be generous enough that this is rare? `[Daniel decision]`
 - **OQ3 — Configurable total in-flight memory cap.** Should there be a single hard byte ceiling on total
  decoded audio held by the player (a safety net independent of the window-size policy), exposed as a
  config value? Recommend **yes, as a guard rail** even if the window policy is time-based — it is the
  backstop that makes "1 GB stream never OOMs" a guarantee rather than a tuning hope. `[Daniel
  decision]`
 - **OQ4 — Apply windowing to all tracks, or only long ones?** A 3-minute Cut decoded whole is ~30–60 MB
  — harmless today. Windowing everything is simpler (one code path) but adds refill machinery to short
  tracks that never needed it. Recommend **window everything** (one path, C6-safe, and short tracks
  simply never hit a refill because they fit inside the forward window) — but Daniel may prefer a
  size threshold. `[Daniel decision]`
 - **OQ5 — Is MSE (Direction C) the real destination?** Not for this phase, but it bears on how much to
  invest here. If DeepDrft will move to compressed delivery (Phase 1.2) and MSE within ~a year, Phase 21
  should be the *minimal* Direction-A change (don't gold-plate machinery MSE would retire). If WAV +
  bespoke graph is the long-term commitment, a more thorough windowing investment is justified.
  `[Daniel steer — informs scope, not a blocker]`
 ---
 ## 7. Acceptance criteria
 - **AC1 (headline) — Bounded memory under a 1 GB stream.** Playing a 1 GB+ WAV mix start to finish, the
  browser tab's retained decoded-audio memory stays bounded to the configured window (not growing toward
  ~2 GB). Verifiable via browser memory tooling: peak decoded-audio footprint is independent of track
  length and tracks the window-size policy, not the file size.
 - **AC2 — Playback-start latency at parity (C2).** First-audio latency for a track is unchanged from
  pre-windowing (within noise). Windowing does not introduce a fetch-then-play stall.
 - **AC3 — Continuous playback, no starvation.** A long mix plays edge to edge with no audible gaps,
  underruns, or stalls under normal network conditions — the forward fill stays ahead of the playhead.
 - **AC4 — Seek-back within the window is instant (UC3).** A short backward seek into retained audio
  produces no network request.
 - **AC5 — Seek (forward, and back past the window) still works (UC2/UC4).** Both resolve via the
  existing Range path with the same behavior the listener sees today; the pre-seek region is evicted, not
  retained.
 - **AC6 — A mid-stream refill failure degrades cleanly (the 1.6 adjacency).** A failed refill fetch
  surfaces a clear user-visible error and leaves the player in a recoverable state (not a wedged
  "playing" with a starved scheduler). It must not silently hang.
 - **AC7 — The Mix visualizer is unaffected (C7).** With the lava visualizer running on a long mix, the
  visualizer renders identically (it reads the preprocessed datum, never the evicted buffers).
 - **AC8 — Single-decoder concurrency invariant holds (C6).** Under rapid seek + refill activity, no
  interleaved `ProcessStreamingChunk` calls corrupt the single JS decoder (the existing drain/cancel
  discipline still governs every fetch).
 ---
 ## 8. Wave decomposition
 Dependency shape: `21.1 → 21.2 → 21.3`, with `21.4` validating the whole. 21.1 is the cold-start
 prerequisite and the load-bearing change; the rest layer on it.
 - **21.1 — Partial eviction in `PlaybackScheduler` (cold-start; the load-bearing change).** Give the
  scheduler the ability to drop already-played buffers and keep its position/index bookkeeping correct
  against a buffer array that no longer begins at absolute time 0 (today `getCurrentPosition`,
  `playFromPosition`, and the scheduling loop all assume `buffers[0]` is the track start). This is the
  hardest correctness work in the phase — the time-anchor math must stay exact through eviction. No
  refill yet; with eviction alone and the forward read loop unchanged, this is provably memory-bounded
  for the *played* region. **Independent of the §6 open questions** — it can begin immediately; the
  window *sizes* (OQ1/OQ3) are parameters fed in later. Settled and cold-start.
 - **21.2 — Back-pressure on the forward read loop (the bound on the *unplayed* region).** Make the C#
  `StreamAudioWithEarlyPlayback` loop stop calling `ReadAsync` when forward decoded lookahead exceeds the
  high-water mark, and resume below low-water. Together with 21.1, this bounds *both* the played and
  unplayed sides — the full memory guarantee (AC1). Must route resume/pause through the existing
  cancellation-safe single-loop discipline (C6). **Depends on 21.1** (eviction must exist so the drained
  region is reclaimed, not merely un-read).
 - **21.3 — Seek-back-past-window refill (close the random-access case).** Wire UC4 — when a backward
  seek lands earlier than the retained tail, refetch via the existing seek-beyond-buffer Range path
  pointed at the earlier offset, and the minimal AC6 refill-failure handling. Mostly **reuse** of the
  landed seek path; the new work is the trigger (window-miss detection) and the clean-failure path.
  **Depends on 21.1 + 21.2** (needs the window boundaries they define).
 - **21.4 — Validation pass against the 1 GB target (acceptance).** Exercise AC1–AC8 against a real 1 GB+
  mix: memory profiling (AC1), latency parity (AC2), edge-to-edge playback (AC3), the seek matrix
  (AC4/AC5), induced refill failure (AC6), visualizer-running (AC7), and rapid-seek concurrency (AC8).
  Largely test/measurement; any break is likely a tuning fix in the 21.1 anchor math or the 21.2
  water-marks. **Depends on 21.1–21.3.**
 ---
 ## 9. Cross-references (read before implementing)
 - Root `CLAUDE.md` "Streaming-first audio playback" / `CONTEXT.md §3.5` — the seam this phase modifies;
  the §2 invariants here restate its contract. Both flag it as the most load-bearing path.
 - `PLAN.md` Phase 4 (landed) / `COMPLETED.md` — the HTTP Range `bytes=X-` primitive this generalizes.
 - `PLAN.md` Phase 1.3 / 1.4 / 1.5 / 1.6 / 1.7 — the deferred decoder/scheduler-seam features; §5 above
  reconciles each.
 - `PLAN.md` Phase 9 — defines the `Mix` medium (single long track), the canonical 1 GB case.
 - `PLAN.md` Phase 10 / `product-notes/phase-10-mix-visualizer-lava-reframe.md` /
  `product-notes/phase-12-waveform-visualizer-generalization.md` — establishes the preprocessed
  per-track high-res waveform datum; the basis for C7 (visualizer does not read live PCM).
 - `DeepDrftPublic/Interop/audio/PlaybackScheduler.ts` — owns the unbounded `buffers: AudioBuffer[]`;
  21.1 lives here.
 - `DeepDrftPublic/Interop/audio/StreamDecoder.ts` — `reinitializeForRangeContinuation`,
  `calculateByteOffset`; the refill substrate.
 - `DeepDrftPublic.Client/Services/StreamingAudioPlayerService.cs` — the C# forward read loop
  (`StreamAudioWithEarlyPlayback`), the seek-beyond-buffer path (`SeekBeyondBuffer`), and the
  cancellation/drain discipline (C6); 21.2/21.3 live here.
 - `DeepDrftPublic.Client/Clients/TrackMediaClient.cs` — the Range-capable media fetch reused by refill.