docs: save parallel wave delivery and final handoff

OUTSTANDING_TASKS.md

@@ -173,6 +173,10 @@ The maintainer directed (2026-08-30) that the repository goes **private for the
 
 ---
 
+### J.1. September 8 control-path review checkpoint
+
+- [ ] **SC-10 follow-through — maintainer review of the current dependency/CI queue and acknowledgement of the #2772 merge error.** Review #2769 (npm minor updates), #2803 (the coherent Vitest 5 trio), #2791 (nightly observer), and #2792 (CI trust contract) at their current verified heads before merge. `ci/policy.v1.json` explicitly includes frontend package manifests/locks and `backend/Directory.Packages.props` as control paths. The coordinator merged #2772 at `d8fc1ceca3e33bef6ef9d315d79a5e996002ae18` without the required maintainer review, despite green CI and independent review; the error is disclosed on #2772 (comment 5577569078) and #2337 (5577569315). No retrospective approval, acknowledgement, rollback decision, or changed authority is inferred. The checked SC-10 row above covers its historical named batch only. Refresh exact heads, bases, CI and remaining review budgets before acting; the three individual Vitest PRs #2770/#2771/#2773 remain open until their replacement is authorized and merged.
+
 ## K. v0.3 residual decision batch (2026-09-05, from the alpha lane's read-only audit of 15 milestone issues) — human actions
 
 Source: the alpha lane audited 15 open v0.3 issues on 2026-09-04 against `main` `61e94f672` to `8c511205d`, each audit adversarially re-verified read-only, and handed the coordinator the gates that need a maintainer answer. Everything agent-decidable from the same audit is already queued on the lanes without a ruling (#2215, #2214, #1307 batch-approve parity, #2007 AC5+AC6, #1968 Legacy-truthful shortcut map, #2009 contrast-guard pins, #1284 view-level regression, #1961 board-store dedupe, #2501 MEDIUM-2). Answer by letter on the issue named in each item, or in one reply here; each audit item says what the answer unblocks and names a proposed default, which is the direction the agents would recommend, not authorization to act while the item is unanswered (D-12 was ruled in-session on 2026-09-05 and is closed). Nothing below is built, closed, moved or re-titled until its letter is recorded here or on the issue (D-12 was ruled in-session on 2026-09-05; D-1 to D-11 and D-13 to D-15 were ruled in-session on 2026-09-06 through the guided walkthrough, each row carrying its letters).

docs/IMPLEMENTATION_MASTERPLAN.md

@@ -1,6 +1,6 @@
 # Taskdeck Implementation Masterplan
 
-Last Updated: 2026-09-07
+Last Updated: 2026-09-08
 <br>
 Planning Horizon: the revival waves in `docs/REVIVAL_PLAN.md` (truth + safety → transcript engine → open-beta launch → generalist expansion [Phase 4, ADR-0046 Accepted]) plus ADR-0051's bounded autonomous backlog lane, then a maintainer checkpoint on beta traction — _(historical: 2026-06-13→2026-07-10 this was the finite archive-pivot waves; before that an open "Next 8 to 12 weeks" release horizon)_
 Companion Active Docs:
@@ -65,6 +65,37 @@ behaviour.
 
 ## Dated delivery and governance updates (newest first)
 
+## Delivery update (2026-09-08, parallel implementation and asynchronous review)
+
+The maintainer authorized concurrent issue delivery, including v0.4/v0.5 foundations,
+with separate Terra review threads. Seven scoped implementation lanes were admitted;
+the coordinator retained canonical documentation, project state, and merge judgment.
+Capture edit-capability polling, Paper read-only keyboard behavior, inspector/dialog
+focus, and the Processing-state API test isolation have landed through PRs #2796,
+#2788, #2799 and #2801. Parent issues with broader acceptance remain open.
+
+CF-24A's first synthetic text corpus and deterministic per-kind scorer landed in
+PR #2804 (21 Python tests plus hosted CI). #2319 remains open for schema-validation
+and processor-attribution follow-through, expanded media, and measured processor
+quality/latency/cost. The initial command executes no processor.
+
+PR #2805 completes #2418's resumable repair for historical Capture divergence
+masked by later disposition timestamps. Version-zero backlog selection, immutable
+asset append and a distinct v2 completion marker preserve truthful pending reads;
+archived mismatches remain outstanding while healthy rows advance. PR #2806 also
+completes #2777's shared frontend instruction parity corrections.
+
+The representation and processing-policy lanes implement their first contract slices;
+runtime persistence, processing queues and runners remain separate acceptance. The
+benchmark lane starts with synthetic text and explicit unavailable metrics. These
+boundaries prevent foundation code from being mistaken for a completed v0.4/v0.5
+processing product.
+
+Control-path dependency PRs require the maintainer's own review under ADR-0066,
+including frontend package manifests/locks and backend Directory.Packages.props.
+The coordinator missed this requirement on merged PR #2772 and disclosed it on that
+PR and #2337. Green CI and independent review do not supply the missing review.
+
 ## Planning update (2026-09-07, transcript prompt v3)
 
 #2211 moves current generation, frontend provenance recognition and packaged acceptance to

docs/STATUS.md

@@ -1,13 +1,22 @@
 # Taskdeck Status (Source of Truth)
 
-Last Updated: 2026-09-07
+Last Updated: 2026-09-08
 
 **Authority.** This file owns *shipped reality* - what is built, verified, and running today.
 **Evidence and authority:** reconcile shipped-state claims against code, tests and current execution evidence. Applicable instructions and `.agent-harness/tier.json` govern authority; this file cannot grant or override it.
 **Direction** (product identity, release themes, open strategy decisions) is owned by [`docs/strategy/PRODUCT_DIRECTION.md`](strategy/PRODUCT_DIRECTION.md).
 **Execution plan** (phases, waves, ship gate, checkpoint) is owned by [`docs/REVIVAL_PLAN.md`](REVIVAL_PLAN.md).
 **History** - delivery records dated before 2026-07-01 - lives in [`docs/archive/status-history/`](archive/status-history/) and is non-authoritative.
 
+September 8 parallel delivery wave (merged work; #2235):
+
+- Historical Capture text divergence masked by later Keep/Archive timestamps is repaired by PR #2805 (#2418). The additive `LegacyReconciliationVersion` column defaults old rows to zero, bounded SQL pages revisit them once, and successful reconciliation appends immutable superseding assets before stamping the current version. The distinct v2 completion marker ignores historical v1 success. Archived mismatches remain outstanding and readable through safe queue fallback while healthy rows progress; list/detail guards do not rely on timestamps alone for unresolved version-zero rows. The full backend gate passed 8,977 tests with five existing skips, followed by clean independent review and hosted CI. No existing private database or large-database timing was exercised. This supersedes the timestamp-only repair description in the original CF-01 record below.
+- Shared frontend instructions now name the actual Claude adapter/import, runtime-specific search fallback and explicit PowerShell working directory (PR #2806, #2777). Documentation, syntax/path checks, independent review and required hosted checks passed; no runtime behavior or authority changed.
+- CF-24A's first benchmark slice is available (PR #2804, #2319): nine synthetic text sources totaling 934 bytes, explicit origin/license/hash/reference metadata, a 16 KiB source-byte budget, and deterministic per-kind precision/recall/F1 scoring from supplied predictions. Twenty-one Python tests and the required hosted checks passed. The command does not run a processor; audio/image/PDF quality, latency, WER and cost are unmeasured. Schema-validation and persisted processor-identity limitations remain tracked on #2319, alongside the larger corpus acceptance. This establishes fixture/scoring tooling, not a release-quality result.
+- Capture polling preserves the server's `canEditSuggestion` decision (PR #2796), reducing interrupted editing while keeping the server authoritative. Paper read-only keyboard behavior and dialog focus are corrected (PRs #2788 and #2799); the wider #1968, #1999 and #2090 acceptance remains open. The Processing-state capture API regression now disables its background triage worker and proves the persisted state before asserting rejection (PR #2801, #2798).
+- These four slices each passed independent Terra review and their required hosted checks at the merged head. Local focused evidence is scoped to each PR; it is not a new full-suite count for current `main` or a packaged/live-provider acceptance claim.
+- Sentry.AspNetCore 6.10.0 and Testcontainers.PostgreSql 4.15.0 landed in PR #2772 with green hosted CI and independent review. **Authority correction:** the coordinator missed the maintainer-own-review requirement for `backend/Directory.Packages.props`, a declared CI control path. The merge is disclosed on #2772 and #2337; no retrospective approval is inferred. New npm/Vitest and CI-control PRs remain subject to that review under [OUTSTANDING_TASKS.md](../OUTSTANDING_TASKS.md).
+
 Transcript prompt identity (#2211, current implementation):
 
 - Current LLM transcript extraction records `llm-triage.v3`, including successful empty extraction, so the capture-reference-date semantics introduced by #2206 have their own provenance identity. The JSON shape remains schema v2; deterministic extraction keeps its existing identity. Historical `llm-triage.v2` payloads, schema and recorded provenance remain readable. The frontend classifier and current packaged acceptance checks move with the producer. This changes attribution, not review-first approval or Apply behavior.