Take the read-back back out of the persist; a write verdict belongs only where evidence crosses the plugin boundary

Its false gated six undo points, so an unverified byte-equality assumption could have silently removed Ctrl-Z for a bank mutation that landed.

src/shell/capture/CLAUDE.md

@@ -57,7 +57,7 @@ detail not covered there:
 - `render_selection` (`shell/capture`) — the transient track selection a selected-tracks render (`&128`) requires, as a stack RAII guard: REAPER prints whatever tracks are selected, so `renderOffline` makes the request's own tracks BE the selection for the render's duration and restores the user's set on every exit path. Engaged ONLY for that source mode, which leaves a stated residual: a `&32` selected-items render still prints whatever ITEMS the user has selected. Live captures are unaffected (that selection is the source), but a recipe replay of a `SelectedItems` capture renders against whatever happens to be selected then — the recipe stores tracks and a range, never item GUIDs, so this guard cannot close it. Filed in `docs/TODO.md`.
 - `render_isolation` (`shell/capture`) — the transient upstream silencing a ranged ITEM render needs, as a stack RAII guard alongside the two above: the selected-tracks source prints everything flowing INTO the track, so each direct folder child's `B_MAINSEND` and each of the track's receives' `B_MUTE` are cut for the render and restored on every exit path. Direct children only — a grandchild reaches the track through the child that owns it. The child-set walk is pure (`core/capture/track_topology`).
 - `capture_orchestrator` (`shell/capture`) — single-capture orchestration + the realtime/insert action bodies (Q-W3 hoist, T4-02): `renderOffline` (one offline render under the scope's FX-bypass guard), `captureAndIndexOne` (render + provenance stamp + bank add + tracking-ledger record, unpersisted), `RunCapture`/`RunCaptureItemAssign`, `RunCaptureRealtimeTrack`/`RunCancelRealtime` (the realtime action bodies — the in-flight state lives in `realtime_lifecycle`), and `RunInsertSelected` (the ONE deliberate exception to capture-never-places).
-- `bake_land` (`shell/capture`) — the EXTENSION's half of the resample chain, the SCAN PASS: scans every open project tab for pending `rsbake_*` requests, lands the ones belonging to the project this session has loaded (via `bake_landing`, below), and refuses the rest with `WrongProject` — one undo point for the batch, each answered over its own key inside the invoking instance's synchronous action call. It owns every ext-state read and write in the chain. The per-key verdict itself is NOT this TU's: it is `core/wire`'s pure `classifyBakeScan`, so this shell only enumerates, reads, and applies — counting every verdict into a `wire::BakeScanTally` as it goes, printing `wire::describeBakeKey` for EVERY enumerated key (the only thing that names which key is whose) plus `wire::describeBakeScan` whenever any key went unanswered or any answer's write was not confirmed, in one `ShowConsoleMsg`. It PROVES every write — answer or stale-clear — by reading the key back (`wire::extStateWriteLanded`, whose home is `core/wire/ext_state_read.h`); an answer that did not land is the one no-answer the tally alone cannot show. That proof is three-valued (`wire::BakeWriteProof`): a read-back that overflowed, or a throw AFTER the `SetProjExtState` call, reports Unknown; a throw BEFORE it reports Rejected, because the write is then known not to have been made. Each key is materialized before any answer is written, so no `SetProjExtState` in this action mutates a set the enumerator is still walking. Answers are held UNENCODED until after the pass's single persist, so a landing the project would not take is answered as a failure rather than as an `Ok` no reload would honour — `wire::bakeLandingAfterPersist` is the ONE route to a `Banked` landing, and no path here (dedup included) may assign that word itself. The undo block is stack RAII (`UndoBlock`). Both loops are guarded: a throw in the scan still writes the answers already prepared, and a throw in the write-back loop still prints the lines already accumulated — no path through this action can end in a silent console. It RENDERS NOTHING — the instrument already did, through its own engine in its own process, which is what makes the baked audio the sound the user approved and what keeps the voice engine out of the extension's link graph.
+- `bake_land` (`shell/capture`) — the EXTENSION's half of the resample chain, the SCAN PASS: scans every open project tab for pending `rsbake_*` requests, lands the ones belonging to the project this session has loaded (via `bake_landing`, below), and refuses the rest with `WrongProject` — one undo point for the batch, each answered over its own key inside the invoking instance's synchronous action call. It owns every ext-state read and write in the chain. The per-key verdict itself is NOT this TU's: it is `core/wire`'s pure `classifyBakeScan`, so this shell only enumerates, reads, and applies — counting every verdict into a `wire::BakeScanTally` as it goes, printing `wire::describeBakeKey` for EVERY enumerated key (the only thing that names which key is whose) plus `wire::describeBakeScan` whenever any key went unanswered or any answer's write was not confirmed, in one `ShowConsoleMsg`. It PROVES every write — answer or stale-clear — by reading the key back (`wire::extStateWriteLanded`, whose home is `core/wire/ext_state_read.h`); an answer that did not land is the one no-answer the tally alone cannot show. That proof is three-valued (`wire::BakeWriteProof`): a read-back that overflowed, or a throw AFTER the `SetProjExtState` call, reports Unknown; a throw BEFORE it reports Rejected, because the write is then known not to have been made. Each key is materialized before any answer is written, so no `SetProjExtState` in this action mutates a set the enumerator is still walking. Answers are held UNENCODED until after the pass's single persist, so a landing whose pass never got its persist through is answered as a failure rather than as an `Ok` no reload would honour — `wire::bakeLandingAfterPersist` is the ONE route to a `Banked` landing, and no path here (dedup included) may assign that word itself. The undo block is stack RAII (`UndoBlock`). Both loops are guarded: a throw in the scan still writes the answers already prepared, and a throw in the write-back loop still prints the lines already accumulated — no path through this action can end in a silent console. It RENDERS NOTHING — the instrument already did, through its own engine in its own process, which is what makes the baked audio the sound the user approved and what keeps the voice engine out of the extension's link graph.
 - `bake_landing` (`shell/capture`) — landing ONE bake request, split off `bake_land` on the one-request / whole-pass seam; touches no REAPER API at all. Non-mutating `prepareLanding` and mutating `commitLanding` sit under separate catches in `attemptLanding` — a throw before anything was written is a clean refusal, a throw after it is reported as possibly partial. Replace-vs-add comes from `tracking::resampleLanding`; a replace keeps the entry's id and slot and never deletes the superseded file. Hash-dedup applies on the add path only, before the disk write, matching `updateSampleInPlace`'s "an in-place refresh is not an insert" — and a dedup hit still rides the pass's persist, because the entry it points at may be one the same pass just added. A refused index withdraws the bytes this call had just written — the self-cleanup carve-out from prune's deletion authority, stated in `prune_fs.cpp`'s header. It never persists: the pass does that once for its whole batch, which is why no landing may report itself as banked.
 - `capture_batch` (`shell/capture`) — the batch-capture family + re-capture-from-source (Q-W3 hoist, T4-02): `RunBatchCaptureItems` (one sample per selected item), `RunBatchCaptureRazor` (one sample per razor area), `RunRecaptureFromSource` (regenerate a provenanced sample from its recorded source's current state, bank-only). Every unit routes through `capture_orchestrator` so every precision invariant holds; persist is batched to one ext-state write per action.
 - `realtime_lifecycle` (`shell/capture`) — the in-flight realtime-capture state machine + globals (Q-W3 hoist): the action starts it, `OnTimer` drives it per tick via `DriveRealtimeCapture` (a single-pointer-test idle fast path — load-bearing hot-path guardrail), `CommitRealtimeResult` lands a finished capture in the bank, `AbortRealtimeCaptureForUnload` tears down cleanly on extension unload.