pmx-canvas 0.8.2 → 0.9.0

package/docs/plans/plan-018-connected-memory-wave-2.md

@@ -0,0 +1,594 @@
+# Wave 2: connected memory
+
+**Status:** Implemented and stabilized with the Wave 3 foundation. Not released.
+
+Build on the local wave-1 stabilization batch. SQLite remains authoritative; Home is a view,
+not a board, and folders are independent of the many-to-many link graph. The exported
+HomeMap/HomeGraph artboards supply the visual model; newer neutral-folder decisions override
+their category hues, synthetic Home and derived levels.
+
+## Contract
+
+- Resolve Markdown `[[Board]]`, `[[Board#Card]]` and optional `|label` aliases on save.
+  Store durable, board-qualified targets separately from the authored text. Preserve unchanged
+  bindings across renames, never guess an ambiguous title, and keep deleted targets visibly broken.
+  Code spans and fenced code do not create links. Unresolved board links offer explicit creation.
+- Extend native edges with an optional target board ID. The source remains on the active board;
+  this does not permit writes on another board. Qualify both ends in the graph, including when
+  two boards contain the same card ID. Preserve edges in snapshots, undo and persistence.
+- Project boards, cards, membership, portals, wiki links, native edges and previous-board chains
+  once on the server. Home Graph, Map, board backlinks and agent graph reads consume it.
+  Summaries count as seen, not full reads. Missing targets are explicit and unlinked boards remain visible.
+- Home Map groups boards in neutral folder regions; human-dragged positions persist in SQLite.
+  Map and Graph use the existing workspace board pins. Graph supports focus and bounded depth,
+  search/type/context filters and direct board/card navigation without switching boards on read.
+- Existing qualified library search stays. The brief discovers linked boards from the same
+  projection; wave 3 adds richer kinds, reasons, downstream review and lifecycle.
+
+## Build and verify
+
+1. Durable links/edge persistence and one projection across HTTP, SDK, MCP and CLI.
+   Verify renamed, duplicate-ID, ambiguous, missing, inactive-board and restart cases.
+2. Home Map/Graph and wiki navigation to the drawings, with responsive controls.
+   Verify human gestures and agent exploration against the same disposable library.
+3. Update public contracts and bundled skills; run unit/client/type/lint/build/browser gates,
+   inspect rendered states at 600, 1024 and 1920 pixels and in all nine themes, then review.
+
+No release or deployment is part of this work. Stabilization changes and unrelated board artifacts
+must be preserved; wave 3 follows this checkpoint as a separate batch.
+
+## Original Wave 2 checkpoint (before the Wave 3 foundation)
+
+- `bun run test`: 1,418 passed; `bun run test:client`: 276 passed.
+- `DISPLAY=:88 bash scripts/run-playwright.sh --headed`: 248 passed on the final bundle,
+  including the 600, 1024 and 1920 px reference projects and all-nine-theme coverage.
+- Build, typecheck and `git diff --check` pass. Lint has no errors; 64 warnings and 12 infos remain.
+- Live human gestures exercised create/edit/Save, wiki creation and navigation, remote-card
+  connection, delete/Undo/Redo, Map drag/reload, Graph discovery/pin/unpin and receipt See change/Undo.
+  Source and rich-editor drafts and selections survive agent relation changes and missing/restored
+  targets; delayed file-backed previews cannot overwrite a subsequently saved draft.
+- Real CLI graph and stdio MCP exploration match the same library without moving the human's
+  active board. Read accounting preserves captured revisions and never acknowledges omitted links.
+- Adversarial review findings were reproduced and fixed; Oracle's final focused follow-up found
+  no blocker. Inspected live screenshots are in `.amp/in/artifacts/w2-final-human/`.
+
+This checks the orb's real local stack, not a released package or a native host installation.
+
+## Combined stabilization
+
+The later foundation exposed reconnect attribution, native save-lineage, board ownership,
+inactive-board scope and catalog replay races. Their fixes are implemented and documented
+in the public contracts. Oracle's final focused follow-up found no remaining race blocker.
+The final unit/client gates pass (1,437/302), and the clean full headed run passes all 266
+cases, including the 600, 1024 and 1920 px reference projects and all-nine-theme coverage.
+Real browser restore/late-ACK and catalog replay probes also pass. The previous full run
+had 264 passes and two failures: its race-test interceptors did not match the new
+board-qualified state URL. Correcting that harness, without a product change, made both
+isolated cases and the subsequent full run pass. Earlier failure evidence is retained.
+`bun run test:coverage` passes with 64.71% lines and 72.10% functions, above the CI floors;
+the fresh-workspace CLI E2E flow passes with headed browser verification. Build, typecheck,
+bundled-skill checks and diff-check pass. Lint has zero errors (64 warnings and 12 infos);
+`CLAUDE.md` and `AGENTS.md` match.
+Live human gestures cover wiki creation/navigation, Map drag/reload, Graph context/details,
+narrow scrolling, daemon reconnect and restored unsaved drafts. Real CLI and stdio MCP
+exploration preserve the active board; lifecycle retain/retire/restore preserves content,
+validation and attributed audit history. Inspected screenshots are in
+`.amp/in/artifacts/w3-final-human/`. Native Copilot/WebKit, native Windows and registry-installed
+execution remain separate unverified surfaces; absent skill mirrors were skipped, not validated.
+
+## Remaining Wave 3 stabilization (2026-10-11)
+
+The remaining relationship and lifecycle work starts from the published foundation, with
+`fb3554bddd0565d94a8747c79b30866f4d1c434b` verified as an ancestor. The owner requested
+adversarial self-review instead of Oracle. This batch is unreleased at 0.8.2. Commit and push
+are authorized after stabilization; release, tagging, publishing and deployment are not.
+
+Reproduced findings and fixes include lost authored provenance, lifecycle-only changes omitted
+from incremental briefs, stale review drafts acquiring new source preconditions, duplicate
+source entries masking an unavailable observation, and a Markdown due footer below the visible
+card. Live browser inspection also found a scrolling relation legend and Fit placing cards under
+the new controls. Each has a failing reproduction and a passing focused check. A full preliminary
+headed run passed 267 cases and failed the packed-consumer selector: its whole-card text matched
+the new source warning as well as the target card. Targeting the card title fixed that harness.
+
+Owner portal feedback caught the preset selector resetting to “Vocabulary…”. The final legend
+keeps all four preset names across reload, reports “Custom vocabulary” after a kind edit, and
+uses the `Uses` drawing's 24 px selector, rounded count chips and one highlight per vocabulary.
+The reset reproduction failed; the corrected headed relation case passes.
+
+The latest serial gates pass 1,443 unit and 302 client tests, typecheck, build and lint (zero
+errors, 64 warnings, 12 infos). Unit coverage is 64.68% lines and 72.16% functions, above the CI
+floors. The final full headed suite passes 268 cases; fresh-workspace CLI E2E passes with
+headed browser verification. Bundled-skill checks pass and the guidance twins match. Agent-skill
+mirrors are absent, so that check skipped rather than validated. The post-debt serial run is
+`/tmp/pmx-w3-post-debt.gM8WHG`; its CLI workspace is `/tmp/pmx-canvas-e2e.IyKnoE`.
+
+Published foundation CI had failures despite the preceding local passes. Review of
+[that run](https://github.com/pskoett/pmx-canvas/actions/runs/38153215292) found Windows subprocess
+imports using URL pathnames instead of filesystem paths, plus test-order contamination: an MCP
+case replaced the shared intent emitter and then disabled it. The local ordered reproduction
+failed in `/tmp/pmx-w3-ci-order-red.FpSrwd`; the same three cases pass in
+`/tmp/pmx-w3-ci-order-green.QSzLU1` after reading the public workbench event stream instead.
+The subprocess imports now use `fileURLToPath`. No assertions were removed and no retries added.
+These are test-harness fixes, not a native Windows UI verification claim.
+
+Live human-style browser gestures cover source editing/Save, relation reason and vocabulary
+editing, preset adoption, category editing, folder inheritance, evidence review, retirement/Undo,
+and explicit recipe handoff. Real stdio MCP and CLI checks cover stale-source rejection,
+qualified briefs and search, retired pinned-resource exclusion, restore preserving validation,
+and relation round-tripping. A second real MCP client reads the human's pinned source and recipe
+steering after restart. Recipe requests do not execute tools or renew validation.
+
+Inspected captures in `.amp/in/artifacts/w3-acceptance/` cover 600/1024/1920 px and all nine themes.
+Failed-run evidence remains in separate `/tmp/pmx-w3-*` directories, including `draft-repro`,
+`duplicate-red`, `legend-red`, `footer-red`, `fit-red`, `preset-red` and the preliminary full run.
+The [maintainability review](plan-009-tech-debt-backlog.md#wave-3-maintainability-review-2026-10-11)
+removed repeated request-local qualification/layout reads. Its measured scaling costs and future
+frame-host, move/audit, journal and runtime boundaries are now checkpoints in the vision.
+These are Chromium/source and local-packed-consumer checks, not native Copilot/WebKit, native
+Windows or registry-install verification. No tag, publish, deploy or version bump is authorized.
+After push, the owner still needs to E2E-test the complete change since 0.8.2 from the updated
+checkout. Running the unchanged registry package at 0.8.2 does not exercise these changes.
+
+## Home acceptance integration (2026-10-12)
+
+The combined batch starts from `0fbc593ac99a7ba34ed6cc59ccc0c49e14ccc606`, with the requested
+foundation verified as an ancestor. The frozen parent source patch was integrated exactly
+(SHA256 `2046aaa21b8e4f326e7da5c99dd7d9042f8ff7802bd871d9dc432b5cc74428f5`), preserving the
+separately owned HomeMemory/layout/live-update changes. Generated bundles, declarations and
+tool catalogs are rebuilt here. The unrelated SDLC artifact is not part of delivery.
+
+Graph now uses live view-local forces, folder outlines and board/card clusters, with counted
+filters, focus/depth, display controls and individual connection reasons. Map preserves saved
+board positions inside collapsible folder regions. Agent edits refresh the projection without
+resetting the human's camera, selection or dragged particles. Board-only controls and native
+paste are unavailable while browsing memory. The parent acceptance fixes add navigation-only
+folder trees, actionable review summaries, safe unused-category deletion, compact settings,
+one-note reviews and ring/hover/legend/Fit corrections. Read, content, validation and relevance
+remain separate clocks.
+
+Adversarial self-review and live transport checks found two additional bugs: the advertised
+MCP board-portal type was rejected by a duplicated enum, and the library context chip omitted
+the pinned active board. Both have red/green regressions, retained in
+`/tmp/pmx-combined-portal-{red.89FSEL,green.OSHgsr}` and
+`/tmp/pmx-combined-chip-{red.jZHY9I,green.805Bqn}`. The earlier library/paste regressions are
+retained in `/tmp/pmx-home-live-red.hjJURq` and `/tmp/pmx-home-paste-red.2qNf8V`.
+The [integration debt review](plan-009-tech-debt-backlog.md#home-acceptance-integration-review-2026-10-12)
+and vision record explicit per-tab editor ownership and measured scaling as future checkpoints.
+
+Live human gestures on disposable data include Graph drag, force controls and pinning, Map
+collapse, Open board, source editing/Save, review confirmation and Not now, and category
+creation/deletion. Real stdio MCP reads the human's board pin, edits a card, and creates portal
+and wiki connections while Graph stays open with the same camera/selection. The observed DOM
+contains both new link kinds and the updated title (`/tmp/pmx-combined-mcp-observed.kO6Ywo`).
+Review evidence records actor human with content/reviewed revision 24. An aborted graph read
+shows Retry; a successful retry restores 114 board/card nodes. The first scratch selector wait
+timed out, but the inspected error capture and the actual Retry click/recovery succeeded.
+
+Inspected renders in `.amp/in/artifacts/combined-acceptance/` include Map/Graph at 600, 1024
+and 1920 px in all nine themes, narrow selected details, collapsed Map, graph failure, category
+scrolling, folder settings and Review. All 54 matrix captures have DPR 2 and no body horizontal
+overflow. Two agent-browser motion recordings failed encoder backlog; their failed files remain
+in `/tmp/pmx-combined-recording-failed*.webm`, not presented as successful video evidence.
+
+This is not full Relations/Flows/Uses sketch parity: the inline edge editor, amber lineage
+strokes, pair/cycle badges and lineage actions remain unimplemented as documented in design.md.
+Dense fit-zoom Graph labels still need selection/local focus for detail; this fixture is not a
+large-library capacity claim. The owner's Amp embedded-portal pill-opening report is unverified,
+distinct from the fixed pin count and from the passing Chromium test-host interaction. Native
+Copilot/WebKit, native Windows UI and registry-installed execution remain unverified. Portal
+creation was exercised through MCP/HTTP, not an expansion of generic CLI/typed-SDK add data.
+No release, tag, publish, deployment or version bump is authorized; the owner still needs to
+test the complete changes since 0.8.2 after the approved commit/push.
+
+Final serial local gates: `bun run test` passes 1,444; `bun run test:client` passes 306;
+coverage passes at 64.71% lines / 72.23% functions. Build, typecheck, bundled-skill checks,
+guidance-twin comparison and diff-check pass. Lint has zero errors, 63 warnings and 12 infos.
+Absent agent-skill mirrors explicitly skip. The demo generator leaves its fixture unchanged.
+`DISPLAY=:88 PMX_PLAYWRIGHT_DB_PATH=<fresh-db> bash scripts/run-playwright.sh --headed`
+passes all 276 cases. `DISPLAY=:88 PMX_E2E_HEADED=1 PMX_CANVAS_E2E_KEEP_WORKDIR=1 bun run test:e2e-cli`
+passes in `/tmp/pmx-canvas-e2e.bQukul`. Logs are in `/tmp/pmx-combined-final-gates.ijVDni`,
+`/tmp/pmx-combined-final-headed.nIX7dJ` and `/tmp/pmx-combined-delivery-gates.VFWQ4K`.
+After the full browser run, only tool-description prose changed; regenerated catalogs/build,
+typecheck/lint and 47 focused MCP/skill checks passed again before the CLI gate.
+
+The additional live two-board vocabulary control sets Platform overview to Dependencies and
+Research index to Flow through the UI. Both retain their own selection after switching and
+reloading; the other 11 boards' vocabularies remain unchanged. Evidence is in
+`/tmp/pmx-combined-vocabulary.SSw0oL` and the inspected vocabulary captures. The full-browser
+context pill opens, but that is not an Amp embedded-portal verification. A scratch Graph click
+ran before Home's asynchronous navigation finished; waiting for Library view before clicking
+completed normally. The inspected final Graph has 114 nodes at DPR 2 with no body horizontal
+overflow; its long sidebar scrolls to the force controls. Board captures retain the existing
+camera, so Fit is still needed when a stored card sits under floating controls.
+
+### Cross-platform startup finding after the combined push
+
+The combined commit was pushed as
+[11dc3117](https://github.com/pskoett/pmx-canvas/commit/11dc311732d8c12acc86dda12b269539df8fa5b3).
+[Its CI run](https://github.com/pskoett/pmx-canvas/actions/runs/38213111457) passed Linux,
+CLI E2E and browser E2E, but both macOS and Windows failed the same concurrent pre-folder
+opening test at `openCanvasDb`'s `PRAGMA journal_mode=WAL`: an immediate `SQLITE_BUSY`
+despite `busy_timeout=5000`. macOS reported 1,443 passes and one failure; Windows reported
+1,422 passes, 21 existing skips and one failure. Full failed job logs remain in
+`/tmp/pmx-combined-ci.fLBUAO/{macos,windows}-failed.log`.
+
+SQLite documents both [busy-handler bypass](https://www.sqlite.org/c3ref/busy_handler.html)
+and [WAL open/close contention](https://www.sqlite.org/wal.html#sometimes_queries_return_sqlite_busy_in_wal_mode).
+The fix retries only journal negotiation within the existing five-second budget, restores
+the usual busy timeout after success, and closes the connection on terminal initialization
+failure. Other errors still propagate immediately; arbitrary content writes are not retried.
+No test assertion, platform skip or CI retry policy is weakened.
+
+Thirty natural Linux repetitions passed before and after, so they are not claimed as a local
+reproduction of the platform race. A disposable fault-injection probe of the exact immediate
+BUSY fails before (`/tmp/pmx-wal-probe-red.QJ9JfP`) and passes after
+(`/tmp/pmx-wal-probe-green.UwfjJ3`). Its negative controls retain a bounded permanent-BUSY
+failure (about five seconds) and immediate corruption failure. The real concurrent/contended
+open regression remains unchanged in the permanent suite; the injected probe is validation
+evidence, not a mock-based replacement for the native-platform gate.
+
+The follow-up headed run also exposed a synchronization race in the file-backed suspended-save
+test. Its trace shows the canonical-content poll arriving at 140508.771 ms, before the test
+registered `waitForResponse` at 140536.895 ms; subsequent incremental polls correctly omitted
+the already-delivered change. The listener now registers before the peer save. The 409 refusal,
+preserved draft, unchanged canonical file and deliberate subsequent-save assertions are intact.
+The failed trace, videos and screenshots remain under
+`/tmp/pmx-wal-final-gates.vWZ0g3/results/boards.pw.ts-save-and-boar-33e76-its-draft-file-backed-true--desktop`.
+
+Final local follow-up gates ran serially: build/typecheck, lint (0 errors, 63 warnings,
+12 infos), 1,444 unit tests, 306 client tests, unit coverage at 64.70% lines / 72.23%
+functions, and fresh-workspace headed CLI E2E all passed. That first browser attempt finished
+with 275 passes and the one listener failure above. After the test correction, typecheck and
+lint passed again, both suspended-save variants passed ten repetitions each (20 cases), and
+the fresh full headed run passed all 276 cases at the configured reference widths. There were
+no browser retries in either local full run. Evidence is retained in
+`/tmp/pmx-wal-final-gates.vWZ0g3`, `/tmp/pmx-wal-corrected-gates.UL7xSE`, and the CLI workspace
+`/tmp/pmx-canvas-e2e.dfySJW`. The client bundle did not change in this follow-up; the combined
+human, real-transport and inspected visual evidence above remains applicable. Native-platform
+CI is still a separate gate, not implied by these Linux results. No release or version change.
+
+### Delivered checkpoint, 2026-10-12
+
+The [stabilization follow-up](https://github.com/pskoett/pmx-canvas/commit/fd63c3c611e680864d444532ca3ef7bb79cd171e)
+is pushed on top of the combined implementation. Its
+[CI run](https://github.com/pskoett/pmx-canvas/actions/runs/38218734543) completed successfully
+across all five jobs: Linux, macOS, Windows, browser E2E and CLI E2E. macOS reports
+1,444 passes and zero failures; Windows reports 1,423 passes, 21 existing skips and zero
+failures. The unchanged concurrent/contended-open regression passes on both platforms.
+Browser CI reports 276 passes with no flaky/retry report. This closes the native-platform
+startup gate above; it does not erase the initial failed runs or their red/green evidence.
+
+A focused read-only Oracle review of the follow-up found no blockers in the startup-only
+retry or save-test synchronization. The owner portal runs this build, with all 11 boards
+preserved; restart checks retain the active board, card contents/geometry, edges and camera.
+The owner's step-by-step acceptance guide now records the delivered build and green CI,
+without marking manual tests passed or replacing the owner's notes.
+
+This is ready for owner acceptance, not a claim of full sketch parity or release approval.
+The embedded Amp pill-opening, dense Fit-label limits, Relations/Flows/Uses gaps and native
+host/UI/registry-install boundaries above remain explicit. Version is still 0.8.2; no release,
+tag, publication, deployment or version bump was performed. Staged maintenance remains in
+the vision and debt backlog.
+
+### Guest cursor polling latency, 2026-10-12
+
+The owner's slow guest-cursor report exposed a PMX transport delay: orb pages select the
+proxy-safe polling path, but that path paused two seconds after every response. Human cursor
+POSTs already throttle at 80 ms and server presence emits coalesce at 16 ms. Neither the
+cursor styling nor the shared board database needed to change.
+
+Direct HTTP clients now request a bounded long poll on the existing event ring. A caught-up
+request waits at most 1000 ms, wakes when an event arrives, and cleans up on timeout, abort
+or server stop. An 80 ms client pause bounds busy-stream request rate. Snapshot/ring recovery
+and SSE are preserved; hosted MCP-app tool polling retains its two-second cadence. No new
+presence channel or host-specific styling was added. Self-review covered event ordering,
+waiter cleanup, client generations and the hosted/direct transport distinction.
+
+The starting checkout contains the requested
+[docs baseline](https://github.com/pskoett/pmx-canvas/commit/175031013ef6bc31b94cec16d5def56661e866b6)
+and the newer docs-only
+[checkpoint](https://github.com/pskoett/pmx-canvas/commit/3b7a570a98bd6f32b8a50386c04bfc81db2a76e6).
+The same two headed Chromium clients moved the pointer through the real orb portal in both
+directions, one workbench inside a cross-origin iframe, on a disposable one-card board.
+Three 100-point sweeps per direction and build produced:
+
+| Observed measure | Unchanged baseline | Fixed build |
+| --- | --- | --- |
+| Remote target updates during movement | 0.47–0.48/s | 11.24–11.54/s |
+| Median target-update gap | 1999–2001 ms | 84 ms |
+| Estimated median rendered lag, including interpolation | 848–934 ms | 213–217 ms |
+| Estimated p95 rendered lag | 1898–2000 ms | 262–266 ms |
+| Receiver animation cadence | about 60 frames/s | about 60 frames/s |
+| Own-cursor count | 0 | 0 |
+
+Lag estimates match rendered cursor x positions to a monotonic input sweep, not a network
+round-trip measurement. The final fixed idle client completed eight polls over six seconds.
+Before/after recordings were inspected. An earlier fixed capture needed dense 100 ms frame
+samples and decoded positions to contradict a coarse automated description of pauses.
+The final-build capture also shows continuous movement rather than the baseline's two-second
+jumps. Two final recording attempts at 60 and 30 fps overloaded the encoder and were retained
+as failures; the inspected recording uses 15 fps. Final recordings are
+`.amp/in/artifacts/cursor-latency/before-final.webm` and `after-final.webm`.
+
+The public HTTP long-poll regression and two-client rendered-coordinate regression both
+failed on unchanged code, then passed with the fix. The existing two-tab grab/lock/yield
+case also passed with SSE forced as a control. Red evidence is retained under
+`/tmp/pmx-cursor-red.RjuzlZ`; focused green build/typecheck, HTTP and headed evidence under
+`/tmp/pmx-cursor-green.GqytFB`. Repeated portal measurements, served-bundle hashes, input and
+render samples and video-frame analysis are in `/tmp/pmx-cursor-comparison.oZBHTD`.
+Earlier failed recordings/navigation and invalid duplicate-tab or throttled measurements
+were retained separately, not counted as successful runs. The final measurement harness
+selects the agent-browser-owned targets, closes extra test tabs and rejects low frame/input
+cadence or duplicate own cursors.
+
+The first full headed run passed 276 cases but timed out in dark HTML refresh continuity.
+Its trace places the stranded poll at 296926.670 ms, between disabling interception at
+296924.232 ms and reenabling it at 296938.877 ms. Ten unchanged focused reruns passed;
+those did not explain or erase the failure. A standalone headed Chromium + Node HTTP
+probe, with no PMX code, reproduced the interception failure: toggling an unrelated route
+stranded fetch 22 before it reached the server (21 requests received and completed).
+Keeping the route installed let completions progress from 234 to 279 over the final
+observation window. The test now installs its route once, preserves every per-revision
+hold, painted-content assertion and screenshot, and uses the existing context cleanup.
+All 12 repeated HTML/JSON-render dark/light cases passed after that harness correction.
+No viewer or product timeout workaround was added. The initial full-run trace/video is
+retained in `/tmp/pmx-cursor-gates.qGwgtE`; the isolated reproduction and unchanged reruns
+are in `/tmp/pmx-cursor-refresh-investigation.SJzGeg`.
+
+The next full headed run passed 276 cases and failed one receipt Undo/Unpin assertion in
+strict mode: bare `[data-node-id]` matched both the card and a transient command-bar chip.
+The card still showed the expected `After B` text. Five card assertions in that case now
+explicitly target `.canvas-node`; the unseen-human-edit and pin-set checks are unchanged.
+Ten consecutive headed repetitions passed. This second failed run, including the DOM trace
+and screenshot, remains in `/tmp/pmx-cursor-final-gates.T08vva`.
+
+The third full headed run passed 276 cases and failed the rich-relation hover handoff
+assertion. The trace shows 184 ms between starting the blank-canvas pointer move and reading
+the class, beyond the intentional 120 ms hover grace; the two requested animation frames
+alone took 115 ms. That test now drives real pointer events with controlled 50 ms crossings
+and a 500 ms blank-canvas control, then resumes normal time. Three repetitions passed,
+removing the grace in a test-only served bundle failed at the short-crossing assertion,
+and three more repetitions passed after removing the mutation. The probe asserted a unique
+bundle replacement and a loaded runtime marker; no mutation code remains in the tests.
+No hover/store/CSS production code changed. The real-time cursor measurements above use no
+controlled clock. The third failed run is retained in `/tmp/pmx-cursor-delivery-gates.ozvSAA`;
+hover red/green evidence is in `/tmp/pmx-cursor-hover-clock.Z1UYwU`.
+
+The fourth full run passed 274 cases and exposed two product races plus a ghost-test sampling
+gap, retained in `/tmp/pmx-cursor-final-browser.7Mdqdq`:
+
+- A painted shortcut overlay could miss immediate Escape while the App keyboard effect still
+  closed over its previous state. The keyboard subscription now updates at layout commit.
+- An earlier same-tab Fit camera echo arrived during a newer Fit animation and cancelled it.
+  Viewport writes now carry the existing tab identity; only the originating tab ignores that
+  echo. Peer tabs and explicit agent commands still apply the camera. This is echo metadata,
+  not an authorization mechanism.
+- The ghost test's best rAF sample was 30.5px from the target, and it did not observe the
+  `animationend` endpoint before removal. It now also samples that endpoint in capture phase.
+  Position/size thresholds and normal frame samples remain; no ghost rendering or timing changed.
+
+The keyboard and camera interleavings both failed deterministically before their fixes. The
+camera case rendered a 250px card instead of completing the newer zoom to 312.5px. Afterwards,
+18 headed checks passed: three each of the two regressions, original Graph and attachment
+cases, ghost settling and two-client cursor motion. The camera regression also checks a peer
+receives both zooms and an agent command still overrides both tabs. Evidence is retained in
+`/tmp/pmx-cursor-races.S1FpCT`.
+
+An intermediate measurement repeated the sweeps after the keyboard/camera fixes.
+Its portal-served bundle hash matched that built bundle; six sweeps and an additional recording
+control maintained about 60 receiver frames/s. Those measurements, hash and failed encodes are
+also in `/tmp/pmx-cursor-races.S1FpCT`.
+
+The next serial gates in `/tmp/pmx-cursor-stabilized.T2t2MW` passed build/typecheck, lint
+(0 errors, existing 63 warnings/12 infos), 1,445 unit tests, 306 client tests, coverage
+(64.69% lines / 72.29% functions, both floors met) and fresh-workspace headed CLI E2E
+(`/tmp/pmx-canvas-e2e.bf3lLB`). The full headed browser gate finished **278 passed, 1 failed**.
+The failure was `canvas-app.pw.ts`'s curated structured notes / AX host-context case, so that
+run did **not** establish stability.
+
+Its trace confirmed a hosted pin-ordering race: poll `since=1378` starts at
+187162 ms, before Clear all's POST at 187168 ms. The poll returns old pin event 1380, restoring
+the cleared pin in the UI; the intended re-pin at 189205 ms consequently submits `nodeIds:[]`.
+Focus reaches the host, but the pinned card is absent. Pin code and hosted two-second polling
+were not changed here; this was not reproduced against a separate baseline build and is not
+claimed conclusively pre-existing. The trace, failure screenshot and video remain with that
+run; the blocker was reported to the parent orb before further changes.
+
+The pin revision fence now covers local writes as well as incoming frames. Poll replies that
+cross a local write or a newer authoritative read skip only their stale pin frames, then read
+the canonical server pin set from the existing event snapshot, without replaying its other
+frames. This costs a snapshot read on pin reconciliation rather than adding a new endpoint.
+The agent-facing pinned-context brief is not authoritative for membership: it filters retired
+cards. Pending writes reconcile after completion, including refused and normalized requests.
+Board changes invalidate old-board continuations. This adds no
+server clock, changes no 20-pin limit and does not suppress later peer or agent writes.
+Both delayed incremental-poll and snapshot regressions failed before the fix
+(`/tmp/pmx-pin-order-red.ahE6cW`), then passed alongside the original hosted-context and restore
+cases (`/tmp/pmx-pin-order-green.K53uol`, four headed passes). Self-review also reproduced and
+fixed an old-board pending-write leak (`/tmp/pmx-pin-board-red.1q5kkz`): a previous board's
+late acknowledgement must not block canonical normalization or release the new board's
+pending Clear all. No test timeout or pin assertion was relaxed.
+
+The following full run (`/tmp/pmx-cursor-pin-gates.h4HfvL`) finished 280 passed / 1 failed:
+the camera regression completed both tabs' zoom and agent-override assertions, then its
+`route.fetch` was closed during fixture teardown. The route now drains with
+`unrouteAll({behavior:'wait'})` before the fixture closes; errors are not suppressed.
+During that run, self-review found the filtered-brief/raw-membership distinction above.
+Its built bundles and pre-correction patch are retained separately from the end-of-run
+source. The retired-pin regression failed against that intermediate bundle
+(`/tmp/pmx-pin-retired-red.Qm0kPq`), so that run is not counted as final-build verification.
+
+**Final serial gates passed** in `/tmp/pmx-cursor-delivery-final.mwP76f`: build/typecheck,
+lint (0 errors, existing 63 warnings/12 infos), 1,445 unit, 307 client, coverage (64.64% lines /
+72.25% functions), fresh-workspace headed CLI E2E (`/tmp/pmx-canvas-e2e.5XOntZ`), 12 repeated
+headed pin/camera regressions and all **282 headed browser cases**, including 600/1024/1920
+reference widths. These execute real HTTP and stdio MCP paths as well as human UI gestures.
+The table's final six portal sweeps, idle control and served hash are in
+`/tmp/pmx-cursor-live-final.gi4axw`; the served index SHA256 matches the final build
+`a5f5a1c1f73d43c3254c79601cc9e303c63d839983e1f569f82f0d45c95b229d`.
+The inspected final clip shows continuous movement, without blanking or two-second stalls.
+A live-check setup attempt used the wrong supervised port and inconsistent browser launch
+flags, leaving blank tabs; it was corrected before these measurements, not counted as a run.
+
+This demonstrates the PMX-side delay and its fix through an orb portal, not the exact Amp
+embedded container. Native Copilot/WebKit, Windows UI and registry installation were not
+exercised for this batch. The owner's live SQLite content was not used or modified. The
+owner subsequently authorized a combined commit/push from the parent orb after validation;
+this orb hands back source and evidence only. No release, tag or version change is authorized.
+
+### Combined acceptance follow-up, 2026-10-12
+
+The parent orb integrated that frozen patch with the owner's card/relation star removal.
+Authorship remains in hints, including bare lines; pins, read marks and the session lens
+are unchanged. Star regressions failed before the removal and passed afterwards. Headed
+600/1024/1440/1920 width and nine-theme inspections retained the hints. Card corner resizing
+persisted after reload in a disposable database; test 01 means resizing the browser or pane,
+not the card. The live acceptance board has a focused agent review card and relation.
+
+The first combined full headed run passed 281 cases and failed attention history. Its trace
+shows pin POST at 231783 ms, an incremental pin frame captured at 231792 ms, and canonical
+snapshot reads at 231796/231817 ms. Revision fencing skipped the older frame correctly but
+canonical reconciliation did not update the semantic attention reducer, so the card stayed
+pinned while its Context updated notice/history disappeared. This is a product regression,
+not a weakened selector or timeout. Feeding only the accepted canonical pin frame into the
+existing attention reducer fixes it without replaying other snapshot frames. The extended
+existing attention test holds the incremental frame: it failed before the fix, then passed
+five consecutive headed repetitions, including later near-pin history. First-failure evidence
+is retained in `/tmp/pmx-combined-acceptance-final`, red proof in `/tmp/pmx-attention-red`,
+and focused green in `/tmp/pmx-attention-green`.
+
+Final combined serial gates passed in `/tmp/pmx-combined-acceptance-corrected`: build,
+typecheck, lint (0 errors, 63 existing warnings/12 infos), 1,445 unit, 307 client,
+coverage (64.67% lines/72.25% functions), fresh headed CLI E2E with real HTTP/stdio MCP,
+and all 282 headed browser cases. Agent-skill mirror validation skipped absent trees;
+that skip is not a validation claim. The preceding 281-pass/1-fail run remains recorded.
+The initial CLI invocation used the wrong headed environment key; its successful result is
+not counted as headed proof. The final invocation uses `PMX_E2E_HEADED=1` explicitly.
+
+### Owner rehearsal fixes and shared Home relation presets, 2026-10-13
+
+The medium rehearsal orb found five numbered failures: 01 toolbar crowding with two sessions;
+07 inert expanded Markdown wiki links; 11 narrow Graph Updates/filter/search overlap;
+14 blank Medium cadence acceptance; and 19 HTML replacement errors displacing healthy paint
+and graph data/spec updates leaving stale charts. Its frozen 15-file patch fixes those
+contracts. It passed 1,445 unit and 307 client cases, headed CLI and 288 headed browser cases
+before the final localized Graph-search CSS correction; nine targeted width/chrome cases
+verify that final worker state. Those counts do not certify the parent's combined changes.
+Intermittent Plex loading and a Flow Start failure passed unchanged focused repetitions and
+the quiet full run, but their causes remain unconfirmed, not diagnosed away.
+
+The owner's clarification moves editable relation templates into Home, shared like Categories.
+Explicit application copies a saved template onto a selected board without rewriting edges;
+later template edits/deletion do not retroactively change existing boards. The board dropdown
+filters All relations plus kinds actually present, not preset names or Custom vocabulary.
+The obsolete per-board editor is removed. The same public registry exposes template CRUD
+through HTTP, SDK, CLI and MCP; existing scope fences protect workspace writes. The catalog
+persists in the workspace database's existing meta store, including deleted starter templates.
+
+The kept public-operation regression failed before the new operations and passed afterwards.
+It verifies restart persistence without changing cards/connections. Existing transport/fence
+tests now cover the two actions. The browser contract covers Home creation, reload, application,
+deletion, two asymmetric boards, unchanged edges and clearing a vanished kind filter.
+Its first run reached all feature assertions but failed test cleanup on an invented DELETE
+edge URL; the trace proves 404. Cleanup now uses the documented DELETE body. This was a test
+error, not a product refusal, and the failed attempt remains in
+`/tmp/pmx-home-relations-gates/first-focused-failure`.
+
+Rendered inspection caught Home's new button clipping at 390 px (right edge 394.17).
+The same executed boundary check failed before toolbar wrapping and passed afterwards
+(right edge 187.61). The browser contract retains that width boundary. Narrow dialog controls
+fit horizontally and remain reachable by scrolling; screenshots are inspected in
+`.amp/in/artifacts/home-relations/`. The owner portal's old bundle stays unchanged while the
+combined tree is built and tested in an isolated local worktree with a separate database.
+
+The main vision, context extension, historical review, design contract, public API references
+and acceptance instructions are aligned with this model. Owner results/notes and the live
+review fixture are preserved, not reset. Owner acceptance, curation evaluation, full sketch
+parity and native-host checks remain separate. No release/tag/publish or version bump.
+
+The first parent combined full headed gate passed 288 cases and failed JSON-viewer recovery
+in srcdoc mode. Its trace shows interception removed at 886609.9 ms, a poll started at
+886612.1 ms without completion, and an accepted recovery PATCH at 886615.6 ms. This matches
+the independently reproduced Chromium/Playwright interception-lifetime stall documented
+above. Both HTML/JSON recovery tests now install one route before navigation and toggle
+replacement failure without removing that route; exact retry counts, healthy-paint and
+recovery assertions remain unchanged. The first failure is retained in
+`/tmp/pmx-home-relations-gates/first-full-failure`, not counted as a pass.
+
+Final combined serial gates pass in `/tmp/pmx-home-relations-gates`: build/typecheck,
+lint (0 errors, existing 63 warnings/12 infos), **1,446 unit**, **307 client**, coverage
+(64.68% lines / 72.31% functions), fresh headed CLI E2E with real HTTP/stdio MCP,
+72 focused server/transport/fence cases, five final focused UI cases, and all **289 headed
+browser cases**, including 600/1024/1920 reference widths. The final full run includes the
+390 px Home wrapping and final Graph-sidebar correction. Main source and the tested isolated
+worktree match byte-for-byte; generated bundles/catalogs are copied from that tested build.
+The preceding CI run passed four jobs but Windows was cancelled. CI for this delivery and
+the owner's requested fresh numbered rehearsal remain separate from these local gates.
+
+### Fresh rehearsal 07: document save before wiki creation, 2026-10-13
+
+The exact `f5032c5` delivery subsequently passed all five jobs in
+[CI 38408653667](https://github.com/pskoett/pmx-canvas/actions/runs/38408653667).
+Fresh manual rehearsal still reproduced a wiki-creation failure twice: expanded Source Save,
+Back to document, click an unresolved link, then Create. Rich-editor blur normalized the
+Markdown and PATCHed revision 23; the dialog's captured pre-normalization content then PATCHed
+the same revision 23. The first write advanced the revision, so the second correctly refused
+binding after creating the board. Waiting in Document did not prevent a later blur save.
+
+The parent regression reproduces that refusal on unchanged product code. The narrow fix
+finishes the expanded document's blur save queue before opening a wiki dialog or navigating.
+It stops if the board incarnation changes or the save leaves a dirty/refused draft. The
+dialog's source capture and server revision guard are unchanged; a real peer write after
+the dialog opens remains refused and its content is preserved.
+
+Local follow-up gates pass in `/tmp/pmx-wiki-blur-gates`: build/typecheck/lint (0 errors,
+63 existing warnings/12 infos), 1,446 unit, 307 client, all 40 boards/Markdown sanitize and
+native Save/ACK/ABA cases, two inline Markdown cases and five repeated wiki cases. The
+unchanged-product red trace and inspected before/after media are in
+`.amp/in/artifacts/wiki-blur/`. A separate headed real-pointer flow created and opened the
+new board; an independent API read confirmed the source's durable board-ID binding. The
+clip holds the last captured frame for two seconds so the actual destination is readable.
+
+At this checkpoint the follow-up is local/uncommitted, with tested index SHA256
+`3c003e08ebd8782c28db46c2cd5c2e777cae1cc23be0afbd9a7138799aa2c233`.
+The owner daemon remains on `f5032c5` while the medium orb completes its full numbered pass.
+Source-only retest handoff precedes combined integration, commit/push and live refresh;
+no current owner results, notes or live database were changed by this isolated validation.
+The medium orb's corrected 390 px evidence establishes the Updates/filter overlap with the
+sidebar **open**, not closed. Two pinned boards and one external writer also push the context
+chip's right edge to 429.09 px in a 390 px viewport. Required 600/1024 controls pass separately.
+
+The extended existing browser regression reproduces both defects: an ordinary filter click
+is intercepted by Updates, and with only the positioning fix the context edge remains
+429.09 px. The local follow-up positions Updates from the measured memory-stage top, so
+wrapped tabs and counts cannot invalidate a fixed header offset. The existing content-based
+compact step now includes external-writer text, preserving avatars, activity access and an
+accessible label. No owner daemon refresh or guide-result reset accompanies these tests.
+First-failure traces and logs are retained in `/tmp/pmx-responsive-gates`.
+
+Final same-source serial gates pass in `/tmp/pmx-responsive-gates`: build/typecheck/lint
+(0 errors, existing 63 warnings/12 infos), **1,446 unit**, **307 client**, coverage
+(64.66% lines / 72.31% functions), and all **289 headed browser cases**, including the
+600/1024/1920 reference projects. The fresh CLI/HTTP/stdio MCP gate also passes with explicit
+`DISPLAY=:88 PMX_E2E_HEADED=1 bash scripts/e2e-cli-coverage.sh`. An earlier invocation omitted
+that flag; its log is retained separately and is not the headed proof.
+
+Inspected final media in `.amp/in/artifacts/narrow-chrome/` shows open/closed populated Graph
+at 390 px, all nine themes, and 600/1024/1920 layouts. Real-pointer recording exercises
+filter open/close, search, context, external activity and Updates. Executed measurements keep
+the context chip inside the viewport (right edge 376 px), and the collapsed Updates trigger
+below Filters and clear of search. Open overlay panels intentionally cover the canvas;
+the narrow graph remains cramped with its sidebar open, and dense Fit labels need local
+focus. These checks do not establish touch, native-host or full sketch-parity acceptance.
+
+The medium orb independently built the frozen five-file source/test patch against clean
+`f5032c5` in a disposable checkout. Its served index matches the parent tested build:
+SHA256 `a42b09b475b9dad7afda0fc68f0cdec49dd43c9bd57d54b381a1efa2525ac2e0`.
+Manual retests pass trailing-newline creation with durable source binding, refusal after a
+genuine peer edit, and populated Graph interactions at 390/600/760/390. Its read-only live
+preservation comparison confirms unchanged layouts for all 11 boards (96 cards, 63 edges),
+including guide results/notes, pins, cameras and tours. Categories, folders and shared presets
+also match. Read inspection advanced `lastReadAt` on three guide cards, not their content or
+evidence validation. No further check depends on the old live bundle. Disposable cleanup and
+the qualified numbered report remain separate; no owner acceptance results are inferred.