Cycle 0052: replays leave, move and open without stale frames, and every Cycle 0051 carried defect is closed #124

Outcome

Every defect that Cycle 0051 carried is closed, by a fix with a test that fails without it or by recorded evidence. This is a bug-fix cycle: it follows #122, which closed with 24 carried defects (its closing comment). It adds no increment: nothing is selected, dropped or reduced (AGENTS.md, Cycle rule).

• Learner. Test and browser replays leave, move and open without stale frames: no leftover CodeLens, no Study page in the Learning summary, no transport from the wrong replay family. Entries and exits behave as the learner chose.
• Agent. A live test run survives a required fact it cannot read. A browser run's refusals say where a stop can go.
• Engineering. The tests stop pinning implementation constants and fixed waits. The doc comments sit on the right functions.

Verified current state. Base: main at bd082540, v0.2.39.0, read on 2026-10-24. Each task cites the code that produces its defect at that revision.

Abbreviations:

• W = extension/src/studyWorkspace.ts
• H = extension/src/ui/studyHtml.ts
• TC = extension/src/tests/studyWorkspaceTestReplayComposition.test.ts
• TO = extension/src/tests/studyWorkspaceTestReplayOpens.test.ts
• TB = extension/src/tests/studyWorkspaceTestReplayBrowser.test.ts
• TS = extension/src/tests/studyWorkspaceTestReplaySource.test.ts
• QA = extension/scripts/study-bootstrap-qa.mjs

Done when:

• All 24 tasks below are checked, each with "fixed in " or "evidence: ".
• The N52 native round passes in both hosts, as described below.
• The 15-check exact-tree gate passes on the final tree, and required CI passes on the exact head.

Tasks

The source of every task is Cycle 0051 (#122), with the label named. A finding made during this cycle becomes one more task. Proof is one of these three:

• A product fix (tasks 1–7, 11–21, and 24 when it finds a group) lands with a test that is red on bd082540 and green after the fix.
• A test-quality task (tasks 8, 10, 22) proves its tests by a temporary mutation of the behavior they protect: each test is red under the mutation, and the PR records it.
• A defect found during this cycle lands with a test that is red on the commit before its fix.

When a red test needs harness code added later in this cycle, only that test support is ported to the base for the red run. A compile failure or a missing fixture is not a red result.

Evidence closes only the tasks whose Closes line allows it (6, 7, 9, 23, 24). Evidence names the observation and where it is recorded.

1. N51-V1: a stale CodeLens during a pending move. Fixed in 75dec02f (no CodeLens, O8) and f0c11ead (the placeholder goes in at line 1); host tests red without each fix; N52 (b) PASS in both hosts on f0c11ead (Cursor take 2 380d57ee…, VS Code take 1 e2166ac6…).

   • Defect: leaveNodeHistoryMoment (W:25003-25006) clears the slot's stop and fires slotLensChanges. The editor still draws the previous stop's label (Recorded: app/lib/actions.ts:35 · 20.56 s) beside the pending target until it asks the provider again: 2 to 3 frames in VS Code, about 9 in Cursor (W:779 notes the 300 ms settle, NODE_HISTORY_LEAVE_SETTLE_MS). In Cursor, a move to an output also keeps the old source's sticky-scroll header for 2 frames.
   • Closes: no frame of a pending move shows the previous stop's label or header. The fix withdraws the label in the same step as the decoration, for example a pending lens whose title names the target, or no lens.

2. N51-V2: the summary is not withdrawn when a test replay retires. Fixed in c4a7f6e (test red on bd082540, green after).

   • Defect: retireNodeHistorySelection (W:12743-12773) never calls clearLearningSummary. leavePlayback (W:12712-12723) returns early for a test replay because recordingActive is false. The summary is cleared only by the later open() (W:8192-8193). Any publish before that open draws the Live Study content in the Learning summary: 1 to 3 frames in N51.
   • Closes: a test replay's retirement withdraws the summary before any publish, as a browser replay's departure does.

3. N51-V3: a test-family transport shows during a departure. Fixed in c4a7f6e (test red on bd082540, green after).

   • Defect: shownReplayFamily() (W:13588-13591) returns test for any summary composition without a recording. leavePlayback sets recordingActive=false while summaryComposition stays true until open() recomputes it, so publishes in that gap (W:12772, W:12815, refresh/renderSelection before W:9170, and Reload at W:8900) send replayFamily:'test' with no history. N51 saw 1 to 2 frames of Test journal playback in the bottom panel on Return to Live from both families.
   • Closes: the family comes from what is shown (a selected or opening test replay), never from the absence of a recording.

4. N51-V4: the Application tab keeps a browser title. Fixed in c4a7f6e (test red on bd082540, green after).

   • Defect: ensureReplayViewport (W:12417) sets codebook · Recorded application. Nothing renames the panel on departure: dropReplayViewport closes the viewport without disposing the panel, and presentNodeHistoryApplication (W:15063-15091) writes only HTML. So the title stays in the Live composition and in a later test replay.
   • Closes: after leaving a browser replay, the Application pane's title is codebook · Application (its creation title, W:18863) in the Live composition and in a test replay. During a live run it stays the live lease's title, codebook · Application <run id, 8 characters>.

5. N51-V5: a live run ends on an unreadable required fact. Fixed in cadafe96. The root cause is recorded in this Issue: imported names live only in Vitest's module bindings. Tests red on 83f9128e: four inspector cases, the owned-run stop and the service refusal (EVIDENCE_UNAVAILABLE, run paused, next continue goes on). N52's live observation follows on pt3's candidate.

   • Defect: inspectNodeTestVariables (utils/inspectNodeTestVariables.ts:62-70, called from ownedNodeTestRun.ts:691) throws Required Node debugger fact is unavailable or too large: value:revalidatePath.mock.calls.length. continueNodeTest (service.ts:1561, :1593-1595) turns it into RUNTIME_FAILED, which ends the run.
   • Closes:
     • The root cause of this path is recorded in this Issue: whether the inspector cannot walk mock → calls → length (properties of a function and an array in the debugger's variables), or whether the value exceeds REPOSITORY_LITERAL_LIMIT_CHARS (1,024).
     • A path the debugger can read is read.
     • A required fact that still cannot be read refuses that codebook_continue call with the existing EVIDENCE_UNAVAILABLE code, and the text names the fact. The run stays paused at the same stop, so the agent can continue without that fact. MCP 1.8's schema is unchanged; only the code and the run's state after the refusal change. Optional facts keep today's omission.

6. N51-V6: a browser replay's open shows the Study document first. Fixed in c4a7f6e (test red on bd082540, green after).

   • Defect: the summary view gets the generic studyHtml('context') document, then a publish inside open() (W:19775-19778, W:19318-19323, W:8542-8549). loadPlayback calls that open (W:12258) before the presented bundle exists, so 1 frame of Study content shows before the replay's summary. Return to Live shows the same frame.
   • Closes: the summary view receives no Study message before its replay's first message, on a browser replay's open and on its Return to Live. This holds whether or not main before Cycle 0051 (8dd675b9) shows the frame. Close by evidence only if the host test shows no such message and N52 shows no such frame.

7. N51-C2: a browser run of a page link found no stop and sealed partial. Both parts closed in 761500dd, with root causes recorded in this Issue.

   • 7a: the interrupted seal. Fixed: the seal names the run's own failure (test red on cadafe96).

   • 7b: the stop outside the launched route's chunks. Evidence of the boundary (a stop is armed before input in the launched route's client code), and the refusal now names the file and the page and says where a stop can go (test red on cadafe96).

   • Defect: a Cursor Agent run of the top page's Log in link to /login was refused. The refusal: Selected source is absent from the selected application route chunks; verify that it belongs to the launched page (utils/readDevelopmentBundle.ts:117-134; configuredBrowserRun.ts:584). It retried with other stops. Its recording sealed with Runtime interrupted or cleanup incomplete; recorded prefix retained. (service.ts:3620-3627), with no application frame and no source.

   • Closes: a root cause, recorded in this Issue, for each of two questions:

     • (a) Why the seal reports an interrupted run when the AI stopped the run itself.
     • (b) Whether a stop in the navigation target's client component (app/ui/login-form.tsx) can be armed once its chunk loads.

     Each part closes by a fix with its test, or by evidence of a documented boundary (the run profile cannot arm a stop outside the launched route's client chunks). If evidence, the refusal sentence is changed to name where a stop can go (the launched route's client chunks), so the agent does not retry blind. That change is a fix with its test.

8. Fixed waits in three host tests (/ship review). Fixed in 30cb0bbe: the timer tests wait for the retired controller to clear its timer and hold no event, and the superseded-move test awaits that move's handling. Mutation proof: a stopTimer that clears nothing, or a retirement that neither pauses nor disposes, turns both timer tests red; a superseded move that reports a refusal turns the third red.

   • Defect: TC:769-770 and TC:799-800 wait setTimeout(resolve, 600), and TC:943-944 waits 100 ms, each marked gstack-timing-ok.
   • Closes: each wait becomes an await on the observable it stands for (the harness's event or state), keeping real production deadlines (AGENTS.md, Commands and checks).

9. Doc comments above the wrong functions (/ship review; CodeRabbit). Evidence: comment-only diff cad14578, lint and typecheck pass.

   • Defect:
     • H:787-791 (playTestReplayControl's JSDoc) sits above playTransportControl's JSDoc; the function at H:814 has none.
     • H:2469-2474 (testReplayTitle) sits above renderFamilyTransport's JSDoc; the function at H:2517 has none.
     • H:2077-2081 (focusSummaryStart) attaches to controlEverFocused (H:2083), not to the function at H:2088.
     • utils/describeNodeTestReplay.ts:38 (Show more and less) sits above the playingStatus doc.
   • Closes: each block stands directly above its own symbol.

10. Test expectations pinned to implementation constants (CodeRabbit). Fixed in 2e8b9b9a: both expectations are literals; changing either constant's text turns its tests red.

    • Defect: TC:195 compares a notice with SUMMARY_PLACEMENT_OFFER, and TC:838 with STUDY_REMOVAL_REFUSALS.nodeHistory; the constants are imported at TC:8-9.
    • Closes: both expectations are hard-coded literals (AGENTS.md, DAMP).

11. An unanswered notice holds the entry turn (/ship review). Fixed in 75dec02f (host test red on d4ecc320).

    • Defect: enterCompositionForEntry awaits the non-modal "Close browser replay" notice inside the turn (W:11697-11707, W:11770-11775). Every later Open replay or Saved studies entry waits on previousTurn until the learner answers.
    • Closes: a later entry is not blocked by an unanswered notice. It supersedes it, or it proceeds, and an answer to the old notice closes nothing (README's existing sentence).

12. Close and Return to Live do not invalidate a waiting entry (/ship review; Codex P2). Fixed in db06b20 (test red on bd082540, green after).

    • Defect: neither close() (W:9446-9480) nor returnCurrentSelection (W:9131-9186) bumps nodeHistoryEntryRequest, and the entry flow takes no lifetime (W:11695-11712, W:11833-11835). After Close, a waiting second Open replay passes its ticket checks, and open() reopens the reader the learner closed.
    • Closes: Close and Return to Live retire every waiting entry, and a retired entry changes no view, focus or selection.

13. A stale-revision Saved studies open leaves an empty summary composition (/ship review). Fixed in db06b20 (test red on bd082540, green after).

    • Defect: the summary composition is entered first (W:11704). A stale message.revision (W:23097) makes historyOpenSuperseded return superseded. openNodeHistory treats that as success (W:12120-12122), so leaveEmptySummaryComposition (only at W:11745 and W:11806) never runs.
    • Closes: a superseded open with no newer replay leaves the summary composition, as a failed open does: the Study pane comes back on Saved studies.

14. Open replay with the summary and the bottom panel both hidden shows nothing (/ship review). Fixed in b37912f (test red on bd082540, green after).

    • Defect: enterSummaryCompositionForReplay (W:11826-11832) returns early when the reader is already in the summary composition, so showLearningSummary and the Observed moments reveal never run. The hidden views stay hidden.
    • Closes: Open replay always ends with the summary shown and focusable. The learner's other layout choices (panel sizes, unrelated groups) are kept.

15. A browser study chosen from a test replay's summary bypasses the editor-area deferral (/ship review). Fixed in b37912f (test red on bd082540, green after).

    • Defect: showSavedStudy defers with studyAwaitingEditorArea while the editor area is hidden (W:24514-24521). showChosenStudy then calls leaveEmptySummaryComposition, then open() and renderSelection, which clears the deferral (W:13734) and shows the panes anyway (W:23302-23303, W:11847-11857).
    • Closes: the deferral holds until the editor area is shown, as from the Study pane.

16. showLearningSummary fails an open when its identity changes (/ship review, Red Team). Fixed in b37912f (test red on bd082540, green after).

    • Defect: assertCurrent (W:19262, W:19276-19279) throws STALE_PRESENTATION when shownReplayIdentity() changes during the awaits of the move, focus and waitUntil. The throw leaves open() (W:8544) and fails that entry.
    • Closes: a summary reveal superseded by a newer selection ends quietly. The newer selection's reveal wins, and no entry reports a failure for it.

17. Every move failure is worded as a debugger failure (/ship review). Fixed in 75dec02f (ReplayDebuggerStartError; host and unit tests red on d4ecc320; slot, Application pane and layout failures do not reach this sentence, they keep their own rows).

    • Defect: describeNodeHistoryMoveFailure (W:24930-24935, thrown at W:24759) says The replay debugger could not start: … for every error from presentNodeHistoryDebuggerOrKeepOpen, whatever its cause.
    • Closes: the sentence follows the cause:
      • A failure of the replay debugger's start keeps today's sentence and recovery clause.
      • Any other failure reads The recorded moment could not be shown: <cause>. The last moment is still shown. Select this stop again to retry. That covers the recorded-source slot, the Application pane and the layout.
      • <cause> is sanitized as today.

18. forgetUnopenedComposition and a Reload race can leave or blank the summary (/ship review). Fixed in b37912f (test red on bd082540, green after).

    • Defect: forgetUnopenedComposition (W:8634-8640, called at W:8585-8586) drops summaryComposition but leaves the summary document. Reload (W:8895-8902) runs leavePlayback(true), then openRecording. recordingActive is false until loadPlayback sets it again (W:12190), so another open() in that gap clears the summary the Reload kept.
    • Closes: a failed first entry withdraws its summary document. A Reload keeps its summary until its own load ends, unless a newer selection supersedes the Reload: another replay, Return to Live or Close. Then the summary follows that selection. An open() that changes no selection (Restore, Arrange, focus) never blanks it.

19. Step marks are bound by SVG sibling order (/ship review). Fixed in 75dec02f (Reader DOM case red on d4ecc320 in red mode).

    • Defect: tagTestReplayDrawing (H:3019-3031; drawing at H:2407-2434) pairs steps with arrows by their order and labels by nextElementSibling, never by an id from the renderer.
    • Closes: the renderer writes each arrow's and label's step id. The marker reads that id. A drawing whose arrows are reordered or interleaved still marks the right step.

20. Show more's expanded state persists across replays (/ship review). Fixed in 75dec02f (Reader DOM case red on d4ecc320 in red mode).

    • Defect: the only state is the button's aria-expanded (H:224; toggle H:3097-3108; read H:3085; called H:4258), and nothing resets it when the replay changes.
    • Closes: every open of a test replay starts with every clamped text collapsed, whether the replay is another one or the same one opened again. Moves inside the open replay keep the learner's expanded texts.

21. nodeHistoryRefusal is not sanitized at publish (/ship review). Fixed in 75dec02f (host test red on d4ecc320).

    • Defect: W:13377-13380 publishes String(error) minus Error: (stored by report, W:25934-25938), without the path cleaning reportTestReplayFailure (W:12037-12047) and describeNodeHistoryMoveFailure (W:24931) apply with sanitizeCause.
    • Closes: the published refusal has the storage folder replaced, as the other replay failures do.

22. Untested state cells (/ship coverage audit). Add the tests; any defect they find becomes a task. Fixed in 26c4878c: two host cases (TB) and two Reader DOM cases. The failed-open cell found task 59. Red proof: a page mutation for each Reader DOM case fails exactly that case under red.

    • Gap:

      • a browser replay × another saved study;
      • a browser replay × a failed open;
      • the browser replay's Replay from start and Pause clicks (playBrowserReplayControl, H:806-813).

    • Today's coverage: QA:1137-1160 clicks only Previous and Next on a browser replay, and the Replay from start cases (QA:10603-10620) are test-family only.

    • Closes: each cell has a passing test asserting today's contract:

      • Another saved study chosen during a browser replay: the browser replay retires (playback left, summary withdrawn) and the chosen study shows in its own composition.
      • A failed open of a browser replay: the load-failure notification, one Output › Codebook entry, the previous reader state kept, and no summary composition left empty.
      • Replay from start: sends a seek to the recording's start, and the readout shows its start time.
      • Pause: sends pause, and the toggle reads Play.

      The carried patch from Cycle 0051's /ship (a private artifact) holds drafts.

23. Two stale comments (/ship review). Evidence: comment-only diff cad14578.

    • Defect:
      • W:11494 says A test replay opens in the Study pane, not through the browser replay's controller. It now opens in the summary composition.
      • studyWorkspaceTestReplayPane.test.ts:114 says the Study pane's group stays the right column, which the summary composition no longer has.
    • Closes: both comments state current behavior.

24. An empty editor group in the composition (/ship review). Evidence: with VS Code's default closeEmptyGroups modeled, the host case leaves no empty group (test in f028933). The empty group the first run showed came from the harness keeping a group emptied by a move; N51 saw none natively in either host. N52 observes it again.

    • Defect: the harness starts with two groups and never removes an emptied one (studyWorkspaceHarness.ts:446-454, :1073-1074, :1138, :1145-1146). So when the Study pane closes in the summary composition (contextPanel.dispose(), W:8201), an empty group would go unnoticed. N51 saw none in either host.
    • Closes: a host case with closeDisposedPanelTabsLater (harness :5885) shows whether the composition leaves an empty group. If it does, fix it with that test. If not, close with the test and N52's observation as evidence.

25. Found in this cycle: the Study's article source tab stands beside the recorded-source tab. Evidence (no defect): README's sentence promises that a test replay's open does not open the article source. An article tab that was already open from the Study composition stays: the summary composition's layout receipts and Restore use it as the source anchor (studyWorkspace.ts, F402). A trial fix that closed it broke 11 Restore and layout-confirmation host tests, so the layout depends on that anchor.

    • Defect: a test replay opened from the Study composition keeps the Study's Captured-overview.mts tab, which moves into the recorded-source tab's group. Seen in the harness and in N51's Cursor frames (actions.ts, Captured-overview.mts, Captured-actions.ts). README promises no Captured-overview tab beside the replay's recorded-source tab.
    • Closes: the replay's open closes the reader's own unpinned, unedited article source tab once, as it does for a retired browser replay's captured source.
    • Proof: a host test that is red before the fix.

26. Found in this cycle: a Retry storage test failed once under the full parallel run. Fixed in 10dd6d83 (test timing): the second run's refused save goes through the write queue after refresh() returns, and under load the refusal came after the Retry ended. This test and its sibling now wait for that refusal first. A 400 ms delayed refusal fails both on 26c4878c (3/3) and neither after; the line-retirement mutation fails both 5/5, before and after.

    • Defect: studyWorkspaceRemoval.test.ts, "a Retry storage / refresh that saved the earlier study keeps the line …", failed once in the full run of the workspace partitions. It passed 3 of 3 alone.
    • Closes: a root cause (a timing dependency in the test or in the code) fixed with its test, or evidence from repeated full runs.

27. Found in this cycle: a Wiki publisher test failed once inside pnpm health's coverage run. Fixed in 550b5ec0 (product). Root cause: Update anyway's start claimed the marker while this window's ended run still turned it into a reservation under the claim, so it refused itself with Another editor window is starting or ending an update of this Wiki. The start now waits for its own window's change and claims again. 24 parallel runs: 9 and 7 failures in 96 before, 0 in 96 after; a host test holding the reservation's write is red without the fix.

    • Defect: in wikiUpdatePublisherFlow.test.ts, "after a VS Code watch ended on a learner's file, Update anyway on the one refusal of the next press starts the update again …" got ['Open LLM Wiki'] where ['Update anyway', 'Open LLM Wiki'] was expected. It happened in the pt1 gate on addf2e20, so health refused to measure. The file passed alone. The test belongs to an area this cycle did not change.
    • Closes: a root cause (a timing dependency in the test or the code) fixed with its test, or evidence from repeated health runs.

28. Found in this cycle (pt1 /ship review, state-machines specialist): a saved study chosen during or after a failed same-study Reload keeps the old replay's composition. Fixed in 40d30aa1.

    • Defect: the Reload mark (recordingReloadStudyId, task 18) holds the summary composition, but openChosenStudy routed the choice by recordingActive alone, and select() rendered a Study item, so the reader stayed in the browser replay's summary composition for a recording that was gone, or drew a Study item under it.
    • Closes: while the mark is set the reader counts as showing its browser replay (browserReplayShown()): a saved-study choice opens the chosen recording as during the replay, and a step selection is refused until the recording has loaded again.
    • Proof: host tests red on 35e0215e.

29. Found in this cycle (pt1 /ship review, Red Team): an agent reads a reader in a same-study Reload as following the run. Fixed in 40d30aa1.

    • Defect: presentationState() reported contentTab and followingCurrentRun from recordingActive alone, so during a Reload (loading or failed) an agent could start a new run without returning to it, while the reader showed the browser replay; the failed load's reason was not reported.
    • Closes: while the mark is set, contentTab reads recording, followingCurrentRun is false and the failed load's reason is reported; the agent's Return runs.
    • Proof: host test red on 35e0215e.

30. Found in this cycle (pt1 /ship review, state-machine cell entry-opening × Return): a Return to Live while a test replay is opening leaves the reader loading. Fixed in 40d30aa1.

    • Defect: the Return bumps historyRequest, but the open's marker (pendingNodeHistory) kept testReplayHoldsComposition() true through the Return, so the Return could not arrange Live; when the held open resumed it ended superseded and, no longer the latest request, left loading true. Seen for the learner's Return and an agent's returnToCurrent, also on 35e0215e.
    • Closes: the Return ends a test replay that is still opening (its marker, loading state and summary) and arranges Live at once; the open then changes nothing.
    • Proof: host tests (learner and agent) red without the fix (loading: true).

31. Found in this cycle (pt1 /ship review pass 2, state-machines): Reload pressed again after a same-study Reload whose load failed is refused. Fixed in 0a8d85de.

    • Defect: the 'reload-recording' action required recordingActive; a Reload whose load failed before the recording was active kept its summary and its Reload button (task 18), which then threw Select a recording first.; a second press while the Reload still loaded threw the same.
    • Closes: Reload works whenever the replay is shown (browserReplayShown()), reloading the recording the replay last asked for (reloadShownRecording()).
    • Proof: host tests red on 40d30aa1.

32. Found in this cycle (pt1 /ship review pass 2, Red Team): Retry and the Live tab ignore a same-study Reload's mark. Fixed in 0a8d85de.

    • Defect: during a loading or failed Reload, Retry storage / refresh ended in renderSelection, drawing a Study selection under the browser replay's summary, and the 'tab' handler's wasRecording read recordingActive, so choosing Live cleared the mark without re-arranging the reader.
    • Closes: Retry loads the recording again as Reload does; choosing Live re-arranges the reader into Live.
    • Proof: host tests red on 40d30aa1.

33. Found in this cycle (pt1 /ship review pass 3, state-machines and Red Team): a runtime release and a study deletion ignore a same-study Reload's mark. Fixed in 5bd1db27.

    • Defect: retire() rendered the Study selection under the replay, and recordingOpenOrOpening() allowed deleting the replay's study (which its Reload and Retry need), each because it read recordingActive alone while a Reload was loading or had failed.
    • Closes: each reads browserReplayShown(): the replay stays shown, and the deletion is refused with the Recording-open sentence until the learner leaves the replay. The Red Team's third item (the refresh's once-per-run Live entry) is evidence, no defect: that entry requires selection.follow, which a recording's open sets false and only Return or the Live tab (both ending the mark first) set true again; a host test of it passed on 0a8d85de too.
    • Proof: host tests red on 0a8d85de.

34. Found in this cycle (pt1 /ship Step 11, Claude adversarial): nine suspicions checked against the code. Evidence (no defect) for seven of them; the first two (an Application pane write and its retitle during a Reload) proved real through Focus source and became task 37: a Reload's mark after a refused openRecording (the reader shows the failed replay with its Reload and Retry, as after a failed load); an entry ending a saved study's wait before it is refused (an explicit entry ends the wait as an explicit display does, task 15, and no browser replay can be shown during a wait); a rejected layout read in a saved-study choice (reported by reportFailedChoice, as showSavedStudy's own read is); Return clearing loading (the Return supersedes every pending load it clears); pruneClosedTabRecord dropping two rebuilt entries of one address (only when no live tab shows that address, so none is left to close); the quiet STALE_PRESENTATION end (the identity changes only for a newer selection, a Return or a Close, each owning the summary); select() refused during a Reload (the intended task 28 refusal, for the command too).

35. Found in this cycle (pt1 /ship Step 11, Codex adversarial and structured review, Claude adversarial F1/F4): six race and recovery defects in the new paths. Fixed in 40ecb542 (L2 by evidence, below).

    • Defect: (L1) a failed or superseded earlier test replay open left the summary composition while a newer entry was entering it, so the newer replay opened without its summary and transport; (L2) a wait for the editor area that ended during its slot close could arrange a reader closed and reopened meanwhile; (L3) a refused close of the waiting slot tabs dropped their record, so no later display closed them; (L4) an entry superseded during its slot close still asked the one-replay-at-a-time notice and held the turn; (L5) a browser recording opened over a test replay whose read failed left the summary its placeholder, so its failure and recovery did not show (regression of task 2); (L6) a second Reload while the first still loaded started no load.
    • Closes: (L1) only the last entry under way leaves an empty composition; (L2) the wait's end re-checks lifetime, choice and a new wait after its close; (L3) a refused close keeps the still-shown tabs recorded; (L4) the entry re-checks its ticket before the notice; (L5) the summary gets the reader's document again (giveSummaryItsDocument()); (L6) the Reload cancels the pending load first.
    • Proof: host tests red on 5bd1db27 for L1, L3, L4, L5, L6. L2 is evidence: two harness scenarios (Close + Open + a new wait, and a newer entry + a new wait, during the held slot close) did not arrange anything on 5bd1db27 either, because Close ends the timer refresh's signal before the stale wait end reaches open(); the re-check stays as a guard for refresh callers that pass no signal.

36. Found in this cycle (local CodeRabbit CLI round 11 on 40ecb542): the views read no recording during a same-study Reload. Fixed in a6a6a5d0.

    • Defect: message() projected recording, model, contentTab, links, canClick and runRefusal from recordingActive alone, so while a Reload loaded or after it failed the views got the browser family with no recording state (stale status, no failure) and the Live tab.
    • Closes: message() reads browserReplayShown() for those fields; also in this round: the shared test fixture saveOtherBrowserStudy, the JSDoc order of clearLearningSummary, and README's Keep Tab Open exception.
    • Proof: host test (held and failed) red on 40ecb542 (contentTab: 'live', recording undefined).

37. Found in this cycle (local CodeRabbit CLI round 12 on e177d0c6; Claude adversarial F2/F3): Focus source during a same-study Reload draws the Study selection in the Application pane. Fixed in e393b9f6.

    • Defect: presentApplication() returned only while recordingActive; Focus source during a loading or failed Reload reached it through renderSelection, overwrote the replay's Application pane with the Study selection and renamed its tab.
    • Closes: presentApplication() returns while browserReplayShown().
    • Proof: host test (held and failed) red on e177d0c6.

38. Found in this cycle (N52 Cursor take 1 on 8e781153): leaving a browser replay paints "Loading recorded position…" in the Application pane. Fixed in 270022c7; closes with N52 (e) on the new candidate. N52 evidence: no Loading recorded position… frame on (e) from a browser replay in Cursor take 3 (d7f5b0c0…) and VS Code take 1 (6737249e…) on fe37cce9.

    • Defect: leavePlayback called playback.leave() before dropReplayViewport(). The controller's hide() invalidated the shown moment, and the reader's invalidate cleared the still-open recorded viewport (clearRecorded, a reset with clear), so the pane drew Loading recorded position… under codebook · Recorded application for 2 frames (67 ms) before the Live pane (take 742da665…, frames 73–74 after the (e2) press at 154.39 s).
    • Closes: a departure releases the recorded viewport before the controller leaves, so the clear reaches no open viewport; a same-study Reload keeps the clear, which reads true there.
    • Proof: host test (TB) red on 8e781153 (['clear', 'close']), plus N52 (e) on the new candidate.

39. Found in this cycle (hosted CodeRabbit on PR v0.2.40.0 fix: replays leave, open and reload without stale frames, and their entries and exits behave as the learner chose (Cycle 0052, PR-1 of 3) #125, studyWorkspace.ts:9157): a Reload refused after it left the replay drops the failed replay. Fixed in 270022c7.

    • Defect: reloadShownRecording left the playback and set the Reload mark before openRecording checked its refusals. From the Reload button action() checks them first, but Retry storage / refresh reads the storage before it reaches the Reload, so a presentation change begun meanwhile (an agent's restore) made openRecording refuse after the replay's failure and reason were gone and no load started.
    • Closes: reloadShownRecording checks openRecording's refusals (a pending presentation change, a study on its way to the Trash) before it changes anything; after them openRecording only awaits the reader's finished initialization.
    • Proof: host test (TC) red on 8e781153 (readiness idle, reason gone).

40. Found in this cycle (hosted CodeRabbit on PR v0.2.40.0 fix: replays leave, open and reload without stale frames, and their entries and exits behave as the learner chose (Cycle 0052, PR-1 of 3) #125, studyWorkspace.ts:11919): Return to Live does not end an Open replay waiting for the work window's focus. Fixed in 270022c7.

    • Defect: enterSummaryCompositionForReplay waited for the focus with no cancellation; a Return or Close bumped the entry ticket, but the wait went on for its 10 s bound, holding the next entry's turn.
    • Closes: each entry ticket has a cancellation that a newer entry, a Return or a Close aborts (supersedeNodeHistoryEntries); the focus wait listens to it, and the overtaken entry ends quietly.
    • Proof: host test (TC) red on 8e781153 (the entry still waiting after 5 s).

41. Found in this cycle (hosted CodeRabbit on PR v0.2.40.0 fix: replays leave, open and reload without stale frames, and their entries and exits behave as the learner chose (Cycle 0052, PR-1 of 3) #125, studyWorkspace.ts:23435): two reads of recordingActive alone during a same-study Reload. (a) fixed in 270022c7; (b) evidence below.

    • Defect: (a) openReplyOwedTo set opensRecording from recordingActive, so a Saved studies choice during a Reload or after it failed got its pane reply while the chosen recording still loaded; (b) shownReplayIdentity names no replay during a Reload.
    • Closes: (a) opensRecording reads browserReplayShown(). (b) evidence, no defect: a summary reveal outlasted by the Reload's load has run the summary's focus command, and the summary holds its document (the Reload keeps it), before the stale check ends it quietly (task 16); what it skips is the visibility wait, a redundant publish and Cursor's once-per-session Move View offer, which the replay's open already made.
    • Proof: host test (TC) red on 8e781153 for (a) (the reply posted while loading).

42. Found in this cycle (hosted CodeRabbit on PR v0.2.40.0 fix: replays leave, open and reload without stale frames, and their entries and exits behave as the learner chose (Cycle 0052, PR-1 of 3) #125, README.md:19): README's Tracking named Cycle 0049 the active cycle. Evidence: docs diff in 270022c7.

    • Closes: evidence: the Tracking list names Cycle 0052 (Cycle 0052: replays leave, move and open without stale frames, and every Cycle 0051 carried defect is closed #124) active and Cycles 0049–0051 closed with their pull requests; a docs-only diff.

43. Found in this cycle (N52 Cursor take 2 on 270022c7): leaving a browser replay redraws its Learning summary in a departing state for one frame. Fixed in fe37cce9; closes with N52 (e) on the new candidate. N52 evidence: the summary goes from the replay's summary straight to its placeholder on (e) from a browser replay in Cursor take 3 and VS Code take 1 on fe37cce9.

    • Defect: playback.leave() publishes while the summary still holds the replay's document. With the viewport released first (task 38), that publish drew The current captured source is not ready in its native pane… and Loading the selected observation… for 1 frame before the placeholder (take ee8c00ab…, frame 73 after the press at 190.32 s). On 8e781153 the same publish reached the summary but drew no visible change.
    • Closes: a departure withdraws the Learning summary together with the recorded viewport before the controller leaves, so the leave's publish reaches no replay summary.
    • Proof: host test (TC) red on 8e781153 and on 270022c7 (two study messages reach the replay's document during the Return), plus N52 (e) on the new candidate.

44. Found in this cycle (hosted CodeRabbit review of 270022c7, outside-diff finding at studyWorkspace.ts:11951): an Open replay entering from Live waits for the work window inside open() without its entry's cancellation. Fixed in fe37cce9.

    • Defect: enterSummaryCompositionForReplay called open() with no signal when it entered a new summary composition, so open()'s wait for the work window before the summary's reveal (reached when the window loses the focus after the Application group was prepared; an entry that starts unfocused is refused at once) ignored a newer entry, a Return or a Close, and a newer Open replay waited up to 10 s for the earlier entry's turn. Task 40 covered only the standing composition's wait.
    • Closes: that branch passes the entry's cancellation to open(); an overtaken entry ends quietly.
    • Proof: host test (TC) red without the fix (the first entry still waiting after 5 s; its layout response is held while the window loses the focus), plus two test changes reported in the PR: the four existing seams of enterSummaryCompositionForReplay in TC forward the cancellation, and the two tests of an agent's Return that supersedes an entering entry assert that the summary holds no reader document instead of the placeholder (an entry superseded before its open() now builds nothing, so the summary is never touched); a mutation that removes the superseded entry's leave makes both fail.

45. Found in this cycle (hosted CodeRabbit review of fe37cce9, outside-diff finding at studyWorkspace.ts:11946): an overtaken Open replay's Learning summary reveal in the standing composition waits for the summary view. Fixed in c89e7dc2.

    • Defect: the standing composition's showLearningSummary got no cancellation, so its wait for the summary view held the next entry's turn up to PRESENTATION_READY_TIMEOUT_MS (5 s).
    • Closes: the reveal gets the entry's cancellation.
    • Proof: host test (TC) red without the fix (the earlier entry still waiting at 2.5 s), green with it.

46. Found in this cycle (N52 VS Code take 1 on fe37cce9): on (e) from a browser replay, the Application tab reads codebook · Recorded application for one frame over the Live content. Evidence (no defect in the code): presentApplication sets the title before it writes the document, in one synchronous call; VS Code painted the webview one frame before the tab strip (frames 58–59 of 6737249e…), while Cursor shows both in the same frame (take 3, frame 32). The end state reads codebook · Application in both hosts.

47. Found in this cycle (pt1 exact-HEAD gate on c89e7dc2, health's coverage run): a Removal test read a queued archive write right after Stop. Fixed in 4ddbb683 (test timing): a save reports 'saved' before it writes the recording's context and leaves the queue; the fixture and two tests now read the queue after that. A 300 ms delayed saveRecordingContext fails 9 of 204 Removal tests on 26c4878c, none after; the line-retirement mutation fails the same two tests before and after.

    • Defect: in studyWorkspaceRemoval.test.ts, "the line “History not saved: …” of a study whose capture failed and whose run this window no longer holds goes when Retry storage / refresh is pressed" read queued: 1 where 0 was expected in its first snapshot, taken right after action('stop') returns (fixture recordRunWithFailedCapture). It passed in the same gate's test lane; neither the test nor the code it covers changed in pt1. Same family as tasks 26 and 27.
    • Closes: a root cause (the fixture reading the state before the Stop's queued write drains, or Stop returning before it) fixed with its test, or evidence from repeated health runs. Scheduled with tasks 26 and 27 (O1, pt3).

48. Found in this cycle (local CodeRabbit round 2 on f0c11ead): the Study pane's line of a refused test replay control still showed the storage path. Fixed in 8a50e907.

    • Defect: task 21 sanitized nodeHistoryRefusal only; message().errors carried String(error) with the learner's storage folder.
    • Closes: that line loses the private roots and keeps its error name.
    • Proof: host test (TC) red without the fix.

49. Found in this cycle (local CodeRabbit round 2 on f0c11ead): a startup bound's closed-debugger error is worded as a start failure. Evidence (no defect).

    • The bound is bounded(..., true) in ReplayDebugger.awaitSession: its expiry is the replay debugger failing to start. Task 17 words that case with the restart clause.
    • After the expiry the owner reads available === false, so the next selection goes through replaceUnavailableReplayDebugger and restarts it. The clause is accurate.
    • A retirement's expiry is awaited outside the wrap and stays a moment failure.
    • Round 3 on 8a50e907 repeated the finding; checked again: expire() sets detached, so available reads false, and presentNodeHistoryDebugger replaces the owner before the next stop. No new finding in round 3.
    • Round 6 on 50408677 (Minor, replayDebugger.ts:775) repeated it again; same evidence: the expiry is the start failing, and the next selection restarts the replaced debugger, so the restart clause holds.

50. Found in this cycle (pt2 /ship review, state-machines specialist): three recorded-source slot state cells were untested. Fixed in f07e44cc (tests only).

    • Gap: slot-placeholder × leave to another moment; slot-loading × leave to another moment; slot-loading × commit of an output. The behavior was correct; no test held it.
    • Closes: two host tests (TS): a move from an output to the result keeps the placeholder with no stop label and no reveal off line 1; after a failed show (F170's loading text), a move to an output replaces the loading text with the placeholder from line 1, with no stop label.
    • Proof: each test is red under a temporary mutation (the placeholder revealed at line 2; a label drawn for a slot without a stop), green without it.

51. Found in this cycle (pt2 /ship review, Red Team and state-machines): four suspicions checked against the code. Evidence (no defect) for each:

    • (a) A closed debugger's expiry worded as a start failure: the same finding as task 49, same evidence.
    • (b) CHANGELOG's 0.2.38.0 entry still names the slot's CodeLens: each entry describes its own version; this version's entry (the pt2 release commit) states that the tab draws no CodeLens, and NODE_HISTORY_LEAVE_SETTLE_MS's comment says the lens is gone.
    • (c) Other Study pane lines (control, archive, …) still carry full local paths: README's Output › Codebook paragraph documents that some messages can still show full local paths; tasks 21 and 48 align the test replay control's two lines with the replay's other failure lines. Sanitizing every line would change unrelated features, an increment outside a bug-fix cycle.
    • (d) Repeated entries stack identical one-replay-at-a-time notices with dead Close buttons: VS Code's NotificationsModel.addNotification closes an equal earlier notification ("Deduplicate", src/vs/workbench/common/notifications.ts; equal = source, message and primary actions, whose ids are _extension_message_handle_<index> per mainThreadMessageService.ts), and the closed notice's promise resolves undefined, so one notice shows and the earlier entry ends quietly (task 11). Cursor runs the same workbench model; not observed natively there.

52. Found in this cycle (pt2 /ship Step 11, Claude adversarial): a refused test replay control's line shows the storage path after a runtime release. Fixed in ec479ab8.

    • Defect: tasks 21 and 48 sanitized the line at publish with the service's roots; retire() clears this.service, so any later publish had no roots and showed String(error) with the learner's storage folder on the transport's status line and the Study pane's line.
    • Closes: report() stores that kind with the roots replaced while the service names them; message() publishes the stored line (the transport's copy without its error name).
    • Proof: host test (TC) red on f07e44cc, green after.

53. Found in this cycle (pt2 /ship Step 11, Claude adversarial): six further suspicions checked against the code. Evidence (no defect) for each:

    • (a) The slot's whole label is no longer reachable without a mouse: O8; the Call Stack thread Recorded test stop <n> · <path>:<line> and the transport readout carry its path, line and time to keyboard and screen-reader users.
    • (b) withoutPrivateRoots strips a nested error name: withoutErrorName removes one anchored leading name only, so Error: TypeError: x stays whole.
    • (c) A start failure and an abort in the same tick: the move was overtaken, and the newer selection words its own outcome; otherwise task 49.
    • (d) A late Close on an overtaken notice after a Return: Return and Close call supersedeNodeHistoryEntries() (tasks 12, 40), and untilAborted drops its listener when the notice settles.
    • (e) The placeholder reveal yanks a learner's scroll: it runs only when the slot's text becomes the placeholder at a commit, where F204's reveal already puts line 1 at the top; the order was checked natively in N52 (b).
    • (f) A repeated Show more open token: it is ++historyRequest, which only grows per reader, and the webview's state ends with the window.

54. Found in this cycle (pt2 /ship Step 11 Claude adversarial pass 2, and the Step 9 delta review): causes worded after a runtime release had no private roots. Fixed in afdf3fac.

    • Defect: the reader keeps serving its history store after retire() clears the service, but its 11 sanitizing sites fell back to { storage: [] } (Save Learning Replay's notice to the home directory alone), so a control refused after the release, or a failure worded then, showed the storage folder or workspace.
    • Closes: privateRoots() gives the attached service's roots, after a release the released service's (remembered in retire()), before any service the roots the reader holds itself; every site uses it.
    • Proof: two host tests (TC) red on ec479ab8, green after.

55. Found in this cycle (pt2 /ship Step 11 Claude adversarial pass 2): six further suspicions. Evidence (no defect) for each:

    • (a) An overtaken notice's Close button does nothing: after a Return or a Close the browser replay it names is no longer shown, an extension cannot withdraw its own notification, and a newer entry's equal notice replaces it (task 51 d); README and task 11 state that an earlier notice's answer closes nothing.
    • (b) Error: Error: x loses a name: the stored line keeps both (withoutPrivateRoots puts back the one withoutErrorName removed), and the transport drops one, as before pt2.
    • (c) Roots that change later: the stored line already reads <storage>, <workspace> or ~; no old path is kept.
    • (d) A hidden slot tab keeps a stop's scroll: a tab that is not visible has no editor to draw a header, and it lays out the placeholder's bytes when shown; N52 (b).
    • (e) A disposal or retirement worded as a start failure: starting rejects only with the launch's failure or the startup bound's expiry; a disposal resolves it, and a retirement is awaited first (retiring).
    • (f) A permanent failure told to retry: the sentence is task 17's specified wording for every failure other than the debugger's start; a source that does not verify shows the 4A placeholder, not a failure.

56. Found in this cycle (local CodeRabbit round 5 on dba70b06, Major + Minor): a failed move says "The last moment is still shown." when the return to it failed too. Fixed in 50408677.

    • Defect: after a failed presentation, returnToCommittedNodeMoment could not show the committed moment (the slot failed to open, or it threw), and the line still claimed it was shown; a return overtaken by a newer selection also reported the old move's refusal.
    • Closes: the return reports returned, superseded or not-shown; the line keeps "The last moment is still shown." only after returned, an overtaken return reports nothing (DS-F), and the recorded-source row tells a failed return. F368's test expectation is restated (its return throws).
    • Proof: host test (TC) red on dba70b06, green after.

57. Found in this cycle (hosted CodeRabbit on PR v0.2.41.0 fix: Cycle 0052 pt2: a test replay's moves, notices and summary page #126, replayDebugger.ts:779): ReplayDebugger.available's JSDoc said only a replay reopen replaces an unavailable owner. Fixed in 9d85bebc (docs).

    • Defect: the doc omitted that an explicit selection or Inspect of a test replay's stop replaces it too (presentNodeHistoryDebugger → replaceUnavailableReplayDebugger), which led reviewers to read the move's restart clause after a startup expiry as wrong (task 49's finding, a fifth time).
    • Closes: available and awaitSession's @throws name the expiry, which selections replace the owner, and that a seek or Play never does. Proof: evidence, a comment-only diff; the branch is covered by studyWorkspaceTestReplayCells.test.ts:807.

58. Found in this cycle (pt2 exact-HEAD gate on 50408677, health's coverage run): a configured-service test timed out under coverage. Fixed in adb6e2d0 (test bound): nine release waits use RUN_RELEASE_BOUND_MS (10 s), not 1 s. A release delayed 1.5 s fails those nine on 26c4878c and none after; a failByRelease that leaves no error fails the task's test 3/3, before and after.

    • Defect: configuredService.test.ts, "a completed run whose release failed after its saved study was removed by hand names the failed write and the removal as its error, and no storage path", hit Timed out waiting for the failed release of the completed run (its waitUntil bound) only inside pnpm health's instrumented run; it passed in the same gate's test lane and in CI. Same family as tasks 26, 27 and 47.
    • Closes: a root cause (a wait bound too short for instrumented runs, or the code) fixed with its test, or evidence from repeated health runs. Scheduled with tasks 26, 27 and 47 (pt3).

59. Found in this cycle (pt3, task 22's failed-open cell): a browser replay whose open failed wrote no Output › Codebook entry. Fixed in 26c4878c.

    • Defect: a test replay's failed open writes an open:test-replay entry for a report; a browser replay's (openLearningReplay) showed its notification only, so task 22's contract ("one Output › Codebook entry") did not hold.
    • Closes: the failure also writes one open:recording entry with the study, the profile and the sanitized cause; the README says so.
    • Proof: host test (TB) red before the fix (outputEntries: []).

60. Found in this cycle (N52 pt3, VS Code): a node-history reference to an unknown study id answered RUNTIME_FAILED with a raw ENOENT. Fixed in aff71ba9.

    • Defect: Copilot Chat sent …-4735-… for …-473e-… in a history's id; codebook://node-history/<id>/learning-material answered "The study operation failed." with the store's ENOENT: lstat '<storage>/node-tests.v1/<id>' and a recovery about owned resources and clicks.
    • Closes: the node-history resources and codebook_explain_node_history refuse an id with no history folder as NOT_FOUND, naming the id and codebook://node-history; README says so.
    • Proof: host test (nodeHistoryResources.test.ts) red without the fix (raw ENOENT for all three calls).

Tasks 61+ continue in this comment (the body is near GitHub's size limit).

Native round N52 (both hosts)

N52 runs in the Space local:codebook-qa on nextjs-dashboard at 1936e41, at the D034 layout (PRODUCT_GOAL.md, Product constraints: an ordinary desktop display, the editor window expanded, default zoom). It uses each PR's candidate VSIX before that PR merges. The Space's OBS setup and preflight are in AGENTS.md (Native QA). Each PR runs only its own checks:

• pt1: (d), (e), the Application title, the empty group.
• pt2: (b).
• pt3: the live checks.

Hosts and chats:

• VS Code: Copilot Chat, GPT-6 Luna.
• Cursor: Agent, Grok 4.6 High (Fast allowed).
• Fallback: the installed Codex Extension after observed credit exhaustion.

Each host gets a new natural /codebook question about a nextjs-dashboard code path that #107, #118 and #122 did not ask, frozen in a comment before submission.

Checks, in each host:

• (b) moves (task 1): from a saved test replay, press Next stop, Previous stop, Next event (output, then result), Previous event, Replay from start, a slider seek, ArrowRight and Home on the slider, and Play. In no frame does a pending move show the previous stop's CodeLens label or the old source's sticky-scroll header.
• (d) browser open (task 6): open a saved browser replay with Open replay (a browser study recorded in that host if none exists). No frame shows Study content in the Learning summary before the replay's summary. The transport has no event buttons.
• (e) Return to Live (tasks 2, 3, 6): run Codebook: Return to Current Pause from a test replay, then from a browser replay. No frame shows Study content in the summary or a test-family transport (Test journal playback). The end state is the Live composition with the Study pane closed and the summary placeholder.
• Application title (task 4): after (e) from the browser replay, the Application tab reads codebook · Application.
• Empty group (task 24): no empty editor group in the test replay composition.
• Live (tasks 5, 7, on pt3):
  • The question reaches a value the earlier run could not read; if it does not, the result is NOT RUN with the reason.
  • A browser run to a page's link reaches a stop in the launched route, or receives the new refusal sentence.

Validity across PRs. If a later PR changes code that an earlier PR's native check covered, that check runs again on the later PR's candidate. The cycle closes on native evidence from the final candidate for every check.

Task 5 natively. The unit and runtime tests are task 5's proof. The native check is an observation: NOT RUN there does not hold the task.

Measurement. OBS Window Capture at 30 fps; the saved file's frame rate is verified. For each transition, every native frame is inspected from the press to 1 s after the last pixel change. One take per transition per host, with the press time logged. A claim covers recorded frames at 30 fps only, not sub-frame timing. A dropped or frozen frame voids that take, which is taken again.

Acceptance criteria

1. Each task is checked with "fixed in " and its test, proven as its kind requires (Tasks, above), or with "evidence: " where its Closes line allows evidence (tasks 6, 7, 9, 23, 24).

2. Each new or changed test states protects=…; fails_when=…; no_claim=…. The card goes in its Value line for a Reader DOM case, as the existing cases do, or in a comment above the test for a host test.

3. N52 passes in both hosts as above, with each take's SHA-256 recorded in this Issue.

4. The 15-check exact-tree gate passes on each PR's final tree:

   • from extension/: pnpm format:check, pnpm typecheck, pnpm lint, pnpm health:static, pnpm dupes, pnpm dead-code, pnpm test;
   • from the repository root, after that build: node extension/scripts/<check>-qa.mjs verify <new empty folder> for each of study-bootstrap, wiki-reader-dom, study-reading-pose, understanding-focus and repository-return;
   • from extension/: pnpm test:package, then pnpm health;
   • from the repository root: node --test .codex/restore-product-context.test.mjs.

   Required CI passes on the exact head, and hosted CodeRabbit APPROVES that head before the merge.

5. Each user-visible change has a CHANGELOG entry and its sentence in extension/README.md.

6. No existing test is weakened. Any changed test, validator, timeout or threshold is reported with its reason in the PR.

Rules for the implementing agent

1. Acceptance criteria are each test's protects / fails_when / no_claim. Never "make the tests pass".
2. Real code and real tests land together.
3. Any change to a test, validator, CI step, timeout, tolerance, threshold, snapshot or suppression is reported on its own line with its reason.
4. A reported command names the command and its observed result.
5. A report is a claim until the driving agent re-executes it.

Rollback, out of scope

Rollback: revert the PR. No stored format or MCP contract changes. MCP stays 1.8.

Out of scope:

• Any new run profile or increment. Package-root Vitest and new profiles stay later candidates (Cycle 0051: a test replay opens in the region-2 composition, with the Learning summary in the chat's place, one transport and 5A #122 Owner decisions).
• The later candidates C1, C5, C8 to C12, E8 and E9.

Owner decisions

• Standing:

  • The owner's goal of 2026-10-24 (Cycle 0050: every defect Cycle 0049 carried is fixed or closed with evidence #118), and the direction of 2026-10-22: devote every resource to the product; never spend money or break the law.
  • The owner cannot answer, so each decision point takes the recommended option, named here.

• O1 (recommended, /spec): delivery in up to three PRs, merged in order. Each passes its own reviews, CI and gate.

  • cycle-0052-pt1: departure and summary lifecycle (tasks 2, 3, 4, 6, 12–16, 18, 24).
  • cycle-0052-pt2: moves and summary page (tasks 1, 11, 17, 19–21).
  • cycle-0052-pt3: live runtime and hygiene (tasks 5, 7, 8, 9, 10, 22, 23).

  Each PR runs its own N52 checks before its merge (Native round, above). Considered: one PR, which would carry 24 tasks into one /ship review limit of three rounds.

• O2 (recommended, /spec): tasks 5 and 7 start with /investigate. Their root cause is recorded in this Issue before code moves. Considered: patching the symptom (catching the error), which could hide a capture gap.

• O3 (recommended, /spec): no /spec --execute agent. The driving agent owns the cycle and runs every native trial. It may hand a bounded implementation part to a subagent under its review; a subagent never moves HEAD.

• O4 (recommended, /spec): task 5's refusal uses the existing EVIDENCE_UNAVAILABLE code and keeps the run paused. MCP 1.8's schema is unchanged. Considered: keeping RUNTIME_FAILED, which ends a run the agent could continue.

• O5 (recommended, pt1 /ship): the gstack upgrade offered at /ship's start waits (Not now). Changing the skill install mid-ship would change the instructions under a running workflow.

• O6 (recommended, pt1 /ship Step 9): every ASK finding is fixed (tasks 28, 29 and the untested state cells get host tests); the two simplification advisories (redundant waiting-record resets in close() and showStudyAwaitingEditorArea) are kept as harmless defensive resets, as the specialist recommended. Low-confidence appendix items (sanitizing the editor's tab-close error, an entry's reveal failure, the Reload mark's study id after another failed open) were checked and recorded in the PR as by design.

• O7 (recommended, pt1 /ship Step 9): a saved-study choice during a same-study Reload opens the chosen recording, and a step selection is refused until the recording has loaded again (task 28). Considered: letting the choice supersede the Reload and return to the Study composition, which would make a failed Reload behave unlike the replay it still shows.

• O8 (recommended, pt2, task 1): the recorded-source slot draws no CodeLens.

  • The editor redraws a provider's lens only after its own refresh delay (at least 250 ms in VS Code, about 300 ms in Cursor), so any lens names the stop a move left beside the next moment, for frames of every move. A pending lens would be just as stale.
  • The end-of-line label and its hover carry the whole label, which also appears in the Call Stack thread and the transport time.
  • The v9 replay region draws no label line above the code (checked on the PNG, 2026-10-25).
  • Considered: keeping the lens (F207's whole label in narrow groups) and accepting the stale frames, which task 1's Closes line forbids.

• O9 (recommended, pt2 /ship Step 8): the plan audit's one UNVERIFIABLE item (task 1's native proof) counts as verified. The driving agent checked N52 (b)'s takes (Cursor take 2 380d57ee…, VS Code take 1 e2166ac6…) and the candidate VSIX 1976136d… against the hashes recorded here.

• O10 (recommended, pt2 /ship Step 9): the simplification advisory is kept as is. withoutPrivateRoots and publishedNodeHistoryRefusal published two forms (with and without the error name) for two surfaces; the shared causeRoots fallback recurs at six other call sites, a file-wide refactor outside a bug-fix cycle. (Task 52's fix later removed publishedNodeHistoryRefusal, so one helper remains.) The RH-1 test removals (the slot's CodeLens tests, F207, F367, F373) are justified by O8's removal of the lens in the same diff, so Gate Integrity has no finding.

• O11 (recommended, pt2 /ship Step 9): the review's repeat after the test-only fix (f07e44cc) reviews that delta. The product code is unchanged since the full review (seven specialists, Red Team, exploratory QA); a whole second pass would add nothing to the product (owner rule of 2026-10-22).

• O12 (recommended, owner rule of 2026-10-23): pt2's /ship review stops at its third fixing cycle with converged:false. Cycle 3 (afdf3fac, task 54) fixed every finding of the last round; no Critical or Major finding stays open, so no Issue is filed. The final delta was reviewed by the cycle-3 delta review; the local CodeRabbit CLI and hosted CodeRabbit (no round limit) review the final head before the merge.

• O13 (recommended, owner rule of 2026-10-22): no third documentation audit for pt2. /ship allows two audits; both ran (c52-pt2-docs-192447, c52-pt2-docs-193203). Task 56 then added one README sentence and one CHANGELOG sentence for its own fix. Local CodeRabbit (the .coderabbit.yaml CHANGELOG → README rule) and hosted CodeRabbit check them; a third audit would add nothing.

• O14+ continue in the same comment.

Next action

Closed. pt1 d4ecc320 (v0.2.40.0), pt2 83f9128e (v0.2.41.0), pt3 83a07385 (v0.2.42.0, PR #127). Nothing is carried; the next cycle is a feature cycle, written by /spec and designed through /autoplan from PRODUCT_GOAL.md and a new gap audit.