PR #354 — transcript navigation

• Keyboard turn navigation: j/k step turns, gg/G first/last, Ctrl/Cmd+G go-to-turn bar, h/l agent lanes (+ visible lane buttons), f focus — in both inspect and scout.
• Focus view: dedicated event page (replaces LargeModal) with navbar, breadcrumbs, loading state, back/home to origin surface, turn chevrons with keyboard-parity follow arm/disarm. Scope: browser and extension log-dir mode.
• Event links: open in a new tab (ctrl/cmd/middle-click) or navigate in place.
• Event-card fix (original bug): inner-tab clicks (Summary/Messages/API) keep the same turn and card-top alignment (root cause: TanStack resizeItem scroll adjustment).
• Fresh sample visit always lands at top (visit-scoped keys); no cross-visit scroll memory.
• Arrow keys are the only prev/next sample stepper; modified arrows (Alt/Ctrl/Shift/Meta) pass through to the browser.
• ?follow=1: shareable URL that arms live-follow at mount (overrides deep-link stand-down); carried into focus links when already following.

Source: dragonstyle's review comment 2026-07-06T15:10:31Z. Re-verified by Peter 2026-07-20 pre-push.

• c.01 ✅ "loading same data formatting is inconsistent... show top navbar with activity bar animating underneath, loading sample data below in small font with animated elipsis" → fixed: navbar + activity bar + animated ellipsis, scout parity.
• c.02 ✅ "When we exit focus mode, we should be sure to return to the transcript at this event" → fixed: back button exits focus view to the originating transcript at the focused turn.
• c.03 ✅ "the home button needs to return to the original parent (log file or samples list)" → fixed: home / prev-next return to the correct origin surface, relative path and focus view preserved. Scope: origin surface preserved; the /samples grid's cross-log ordering at log boundaries is not (Sol P1-9).
• c.04 ✅ "goto dialog needs additional padding... for long turn counts, the prefill value can be truncated" → fixed: padded, auto-sizing input, floating non-blocking bar (not a modal, per 2026-07-03 decision).
• c.05 ✅ "left side of focus view only shows 'main' rather than the path to the subagent" → fixed: breadcrumb shows full subagent path (useFocusLaneScope.ts).
• c.06 ✅ "turn counts seem sometimes inconsistent... turn 1/12 and then... focus view shows turn 14/25" → fixed: transcript view, focus view, and outline consistent (verified 14/14/14 on petri audit log).
• c.07 ✅ "push the padding / margin out to right/left align elements" → fixed.

Two inline Charles comments re-checked, both still satisfied: getEventFocusUrl in shared url.ts since the original focus-page commit; FocusTabContext already refactored to the callback pattern.

Source: dragonstyle's CHANGES_REQUESTED review 2026-07-21T22:10Z ("this is looking so good! A couple more nits"). 2026-07-22: origin/main merged into the branch (140481c9; textual conflicts trivial; semantic risk of main's hidden-turn changes: see the hidden-turns section below). Fixes implemented and reviewed 2026-07-22 as commits db8f25ad / 652bd671 / d2df5cf3 / 1b70adbb (the 4th folds the utility-toggle parity fix together with the deep-link force-include for both ?event= surfaces — le.48; kept in this PR deliberately rather than split to a follow-up, to be named explicitly in the PR comment for Charles's re-review), unit-tested with red-first guards, Peter-verified per t.65–t.68 and le.48 (t.68 carries the standing qualifier "verified on synthetic data only" — see t.68).

• c.08 ❌✅ outline turn entry (e.g. "9 turns") flashes bold then unbolds during j/k navigation; spec: entry stays bold whenever the current position is within its scope of turns. A race — sometimes ends correctly bold; Peter's 2026-07-20 "looks OK" had too few repetitions to hit it. Root cause found: collapsed "N turns" outline rows are keyed on the FIRST turn's model-event id while j/k and ?event= write the current turn's id, so mid-group selections matched no row; fixed by pure render-time resolveOutlineSelection() mapping any id to its enclosing outline row (ordering-independent, no settle windows); red-first unit test (failed expected 'n8' to be 'n6' on pre-fix semantics). Commit d2df5cf3. Peter-verified 2026-07-22 (t.65): re-verified green across many j/k repetitions, on both the original audit log and the synthetic clienttl copy, closing the sufficient-repetitions reservation.
• c.09 ❌✅ next/prev agent buttons in the swim-lane breadcrumbs area unclear ("not clear what they control"), Charles votes them out of the focus view too → Peter's ruling: remove the visible buttons everywhere, keep h/l keyboard lane hops working as an undocumented power-user shortcut. Buttons removed from swimlane breadcrumbs AND focus view incl. dead plumbing (laneNav, hasPrevAgent/hasNextAgent); h/l keyboard path verified separate. Commit 652bd671. Peter-verified 2026-07-22 (t.66): removal confirmed. PLUS ruling: the h/l regression tests originally added alongside this were "useless slop" — deleted in the slop-sweep rewrite; commit is now production-only, no tests.
• c.10 ❌✅ agent switcher should be a pill matching the full transcript's pill styling — shared styling source, not copy-pasted CSS values. Reuses TimelineSelector's .selectorButton/.chevron directly (cross-file CSS composes rejected — no repo precedent); closed-state button only, dropdown menu untouched. NOTE for retest: intentional size/caret change. Commit db8f25ad. Peter-verified 2026-07-22 (t.67): "looks good, code diff too."

Source: dragonstyle's follow-up review (agent-generated, comment 5051900334 on PR #354) found the branch conflicting with main and a list of issues below. origin/main (chunked-log viewer #451) merged as c08fbb48 — one visible conflict resolved preserving both sides, one silently auto-merged import break caught by gates and fixed in the merge commit; eight fix commits appended, all four gates green (sole failure = known load-flake, 3× green isolated; a second load-flake in the logs-listing fetch engine documented the same way).

• Fable #9 🧪 VirtualList uncancelled initial-scroll rAF (orphan settle loop / direct scrollTop writes after unmount or key change) → fixed, commit 38eb6616; investigation corrected the reviewer's harm path: post-unmount jumps are neutralized by the virtualizer nulling its element; the real damage is direct scrollTop writes and the mounted key-change case.
• Fable #10 🧪 cross-app duplication (chrome nav-ownership block ×4 sites already diverging; event-nav handlers ×4 sites) → consolidated into shared hooks, commits 7d82d1ce + 55724231, per Peter's explicit worry (divergences ruled bug-or-intent from evidence, not averaged).
• Fable #11 🧪 e2e unconditional sleeps → condition polls, commit bec2d634.
• Fable #12 🧪 useScrollTrack re-subscribing on consumer renders → keyed on content, commit b2662882.
• Fable #13 🧪 go-to-turn into a user-collapsed region landed short silently → batch-expands like scrollToEvent, commit a8b82eb6.
• Fable #14 🧪 small fixes batch (selectedTab return type; url.test imports the real directoryRelativeUrl; dead export removed; focus pages set document title; verified perf nits) → commit 9cfd783c.
• Fable #15 🧪 process-vocabulary and oversized comments in code, em-dash-only hunks → cleaned, commit ca387d8e.

REPLIES:

• 🙈 Esc-closes-focus-while-bar-open — standing ruling from round 3 (fixes are risk; reopenable).
• 🙈 plain-arrow capture — arrows-only stepper is the ruled design, modified arrows pass through.
• 🙈 a11y listbox roles — declined as polish per round-3 ruling, tracked as loose end.
• 🙈 missing hook tests — declined per slop-test ruling.
• 🙈 ?event= encoding — pre-existing entrenched pattern, loose-end candidate.
• 🙈 the reviewer itself checked and REJECTED one automated claim (j-during-?message=-settle fight — cleanup cancels the loop; worst case one-frame wobble) — recorded so it never resurfaces.

Review verdict 2026-07-23: CLEAN, no must-fix (independent pass + grok concurrence on the two riskiest diffs). Red-first capability verified for every behavior fix; consolidation confirmed behavior-preserving except the one intended fix (ResultBody now suppresses chrome collapse while find is open, matching the transcript page — the divergence ruled a bug). Two informational notes: a negligible null-scroller expand micro-divergence in ResultBody (harmless, aligns with scout's existing semantics), and the small-fixes commit carries two untested low-risk touches (jumpToEnd scrollHeight target, document titles). Remaining e2e waits are all constants-verified justified-keeps, not polls.

Context for Charles on commit 1b70adbb, the 4th append-only commit on the branch (on top of d2df5cf3), added after merging main's utility-agent classification (#425: "N utility agents hidden" header button) and solver-span keep (#443).

Mechanism. The utility-agent classifier only runs when the client builds a sample's timeline from raw events (classifyUtilityAgents, inside buildTimeline). Samples that ship a server-baked timeline skip it entirely — convertServerTimeline copies utility flags verbatim from the server payload, so a log baked without those flags renders with no "hidden" button and always will, regardless of client code; the audit log used for testing is one such log. Bringing the feature to those logs needs the Python timeline writer to add the same classification at write time; that's out of scope for this PR and tracked as a loose end (le.44).

The bug this commit fixes. The focus page and the h/l lane stepper hardcoded the utility toggle to off, while the transcript view honored the user's persisted toggle. Once main's utility classification made utility lanes selectable in the transcript, selecting one and pressing f (or opening its focus URL) landed on a blank focus page, and h/l in focus covered a different set of lanes than the transcript did. This commit makes focus and h/l read the same persisted toggle the transcript already uses, so a selected utility lane now renders its turn and the lane set matches everywhere. Red-first unit-tested through the real wrapUtilityEvents pipeline (pre-fix failures reproduced verbatim, e.g. lane count off by one; toggle-OFF assertions passed pre-fix, confirming no regression). Peter-verified 2026-07-22 in browser (t.68), on a timelines-stripped copy of an audit log (original left byte-identical, sha256-verified) — standing qualifier: verified on synthetic data only, since no real log in the corpus currently exercises this path (real logs with server-baked timelines never carry utility flags to begin with; see mechanism above). Mirror behavior confirmed: a selected utility lane's turn is reachable in focus view. Not verified/possible: stepping between same-named sibling utility wrappers from within focus view — that is the known lane-model follow-up (le.47), not a defect of this PR.

Deliberately not fixed in this PR:

• A shared deep link (?event=) to a utility-wrapped event, opened by a recipient whose toggle is off, still resolves to nothing silently. Ruled but not implemented in this PR — see le.48.
• Pre-existing on main, worsened by this branch: after toggling utility off, an orphaned selection falls back to the raw event stream and inflates outline/turn counts until clicked away — upstream candidate, see le.44/le.45.
• Swimlane rows that aggregate same-named sibling agent spans (e.g. several utility wrappers in one row) expose only one keyboard stop per row, leaving most instances keyboard-unreachable. A lane-model change (per-instance stops, scoring as its own stop) is sized for a follow-up PR — see le.47.

Charles's second CHANGES_REQUESTED review relayed an 8-item agentic-review list. Verdicts with file:line evidence per item: #1–#7 fixed, #8 not reproducible.

• Fable #1 🧪 VirtualList settle-guard latch on find → FIXED, guard owned by settleScrollToIndex, every exit calls finish() (VirtualList.tsx:674-711, :405-435).
• Fable #2 🧪 FocusTurnView no scroll-reset on turn nav → FIXED, resets scroll per turn and tab (FocusTurnView.tsx:134-136, :152).
• Fable #3 🧪 keyboard capture leak into preview surfaces → FIXED, capture handlers bail on disconnected/hidden containers (useTranscriptKeyboardNavigation.ts:65-71).
• Fable #4 🧪 ?message= deep links land behind the pinned event header → FIXED, sticky-header offset restored via stuckHeaderHeight() (TranscriptViewNodes.tsx:412-440, settle budget 90 frames at :447); landing loop also fixed to abort on user input. Residuals verified 2026-07-20: mid-landing chrome-height growth can end the re-park window early; a settle-cancellation path can skip parking. Parking itself Peter-verified working in practice.
• Fable #5 🧪✅ useScrollTrack.ts:67 binary top-crossing check makes any event in the final viewport unreachable as "current" except the last → FIXED, advancing detection point restored (useScrollTrack.ts:14-24); j/k steps by index instead (t.43); 3 VirtualList race fixes + regression tests restored. Peter-verified 2026-07-20: outline follows manual scroll correctly.
• Fable #6 🧪✅ TranscriptViewNodes.tsx:564 first j skips turn 1 when pre-turn content is on screen → FIXED via scroll-history latch (mount reports ignored until the user actually scrolls; folded into the components slice). Peter-verified 2026-07-20 on both log shapes: petri audit and hcast react. Scope: verified for the load case (two log shapes); scrolling within a tall first turn before the first j still misdirects (Sol P1-6), unfixed.
• Fable #7 🧪 GoToTurnBar.tsx:94 Ctrl+G editable-target check misses shadow-DOM <vscode-textarea>/<select> → FIXED, shared isEditableTarget pierces shadow DOM (GoToTurnBar.tsx:99-104; util/browser.ts:38-62). Overlaps le.28.
• Fable #8 🧪 useFocusTurnNavigation.ts:190 breadcrumb can navigate to the wrong agent instance when an earlier same-named sibling lane has direct turns → NOT REPRODUCIBLE, crumbs resolve by row key + agent span id (useFocusLaneScope.ts:109-118; useFocusTurnNavigation.ts:292).

Two blind reviews (gpt-5.6-sol, grok-4.5) ran against tip c2ccb288 with the report withheld; every finding adversarially verified; only fixes for P1-2 and P1-5 shipped per Peter's ruling ("fixes are just a risk of new bugs").

• Sol P1-1 🙈 arrow-capture steals keys from tree/lightbox widgets → pre-existing on main (branch strictly improved it with the modifier guard).
• Sol P1-2 🧪 focus mode non-functional in single-file mode (one .eval opened directly; AppLayout short-circuits routing — three of four dispatch points have the focus check, AppLayout doesn't) → verified; browser + extension log-dir mode work (Peter-verified); fixed: the missing focus check added to AppLayout's single-file branch, mirroring the three sibling routers (additive); no dispatch-test harness exists to pin it (flagged gap); manual single-file verification pending.
• Sol P1-3 ❌ settle-cancellation can skip ?message= parking → verified residual; folded into Fable #4's scope note (see above).
• Sol P1-4 ❌ live ?message= into the Messages tab can arm follow (navOwned not forwarded by ChatViewVirtualList) → verified; unfixed by ruling.
• Sol P1-5 ✅ k doesn't disarm live-follow (view returns to tail on next event) → verified; fixed: k disarms live-follow before stepping, mirroring the j-past-last arm and the focus stepTurn semantics; red/green unit test added; Peter-verified 2026-07-20 ("k works to stop following just fine").
• Sol P1-6 ❌ scroll within a tall first turn then j steps back to turn 1 (first-j latch hole) → verified; folded into Fable #6's scope note.
• Sol P1-7 ❌ go-to-turn bar: Tab to Close then Escape exits focus mode instead of closing the bar; j/k pass behind the open bar → verified; unfixed by ruling.
• Sol P1-8 ❌ scout: return-to-main from a subagent lane can stay scoped to the old lane (double URL write, stale params) → verified, highest severity of the set; unfixed by ruling.
• Sol P1-9 ❌ focus prev/next under /samples ignores the grid's cross-log order at log boundaries → verified; folded into c.03's scope note.
• Sol P1-10, P1-13, P1-14 already tracked as le.18, le.26, le.31 respectively.
• Sol P1-11 🙈 focus-URL id escaping → unreachable with real ids (severity false positive; defensive nit).
• Sol P1-12 🙈 parking listeners retained after settle window → verified but negligible (passive, single mount).
• Sol P1-15 🙈 scrollPaddingStart divided by scale on >16M-px logs → verified, cosmetic.
• Grok G1 ❌ inspect chrome collapse persists across sample hops (main expands via its live sticky gate; branch seeds ownership once) → verified divergence from main; unfixed by ruling.
• Grok G2 = Sol P1-3 family (mid-landing chrome growth ends the re-park window); folded into Fable #4's scope note.
• Grok G3 🙈 missing aria-labels on scout chevrons → declined ("slop").
• Grok G4 🙈 main-transcript chevron can't arm follow like j → declined ("discoverability affordance; real users scroll").

• t.01 ✅ Samples-list load speed after upstream merge
• t.02 ❌🙈 Logs-list empty-flash on return, pre-existing upstream (#389)
• t.03 ❌✅ Back button in Cursor extension leaving sample view (stale dist)
• t.04 ❌✅✅ f not closing focus view in extension (stale dist)
• t.05 ❌✅✅✅ f losing selected tab entering focus view
• t.06 ❌❌❌❌✅✅✅ Return-from-focus / deep-link landing scroll position
• t.07 ❌❌✅ Summary-header state on landings (expand/collapse blink)
• t.08 ❌✅ f above turn 1 not clamping to turn 1
• t.09 ✅ j/k turn-stepping semantics
• t.10 ❌❌❌✅ Swimlane autoexpand randomness / animation on focus close
• t.11 ❌❌✅ Focus-view breadcrumbs (double "main"; missing branch label)
• t.12 ❌❌✅✅✅ Focus-view nested indent/padding
• t.13 ✅ Charles-reported PR #354 issues, sweep retest
• t.14 ❌🙈 Bare-load chrome on triframe log, artifact overclaimed
• t.15 ✅ Jul-11 checklist sweep (remaining items)
• t.16 🙈 Extended-find during deep-link settle, declined/deferred
• t.17 ❌✅✅ Tab-flip losing scroll position
• t.18 ❌✅✅ Return-from-focus race (turn 1/12 or judge-lane 1/1)
• t.19 ❌✅✅ Live-follow pin to bottom on running eval
• t.20 ❌🙈 Cursor webview yanked to newest log on background reload, deferred to le.01
• t.21 ❌✅ Scroll position leaking between different evals
• t.22 ✅ Error surfacing in focus view (red strip + error card)
• t.23 ❌🙈 STATE UPDATED missing from focus view (debug filter, not-a-bug)
• t.24 ❌✅✅ SCORE card missing from focus view
• t.25 ✅ No red strip / error card on retried-then-green sample
• t.26 ❌❌🙈 Retry eval-set live surfaces (no auto-switch; missing from grid), pre-existing
• t.27 ❌🙈 Background-tab finish-scroll divergence, pre-existing
• t.33 🙈 Focus-view score render vs. transcript-tab substructure parity
• t.35 ❌❌🙈 Shift+L lane navigation across multiple samples in one eval file — superseded by the arrows-only sample-stepper ruling (2026-07-18, t.53): Shift+H/L was removed and modified arrows pass through to the browser. No Shift+L binding exists on the branch tip (verified 2026-07-22); nothing left to retest.
• t.39 ✅✅ Turn-nav keyboard shortcuts (heading-parity complaint spun off into t.41)
• t.40 🙈 Fold-process question (git reset / main branch), not a functional test
• t.41 ❌❌✅✅ Scout focus-view refactor / one-bar header; header-obscured-when-scrollable regression found and fixed
• t.42 ✅✅✅ gg / Ctrl+G chrome-ownership + ?message= deep-link parity
• t.43 🙈 Anti-oscillation j/k opt-out — design decision, not a retest item (see Fable #5)
• t.44 ✅✅ Shift+H/L on scan-result tabs
• t.45 ❌✅⌛ ?message= deep-link event on a slow-loading sample
• t.46 ✅✅ CSS/visual match to main for the chrome-ownership header fix
• t.47 ✅✅ Held-k long-press navigation
• t.48 ❌❌🙈 Event-card expand/collapse + scroll-preservation flakiness, ruled out-of-scope
• t.49 ❌🙈 PageUp/Home releasing chrome ownership, non-bug (identical no-op on main)
• t.50 🙈 Finish-timer yank: running sample finishing within ~100ms of scroll-away pulled the view back to top (stale finish-timer); agent-fixed + unit-tested; same-day regression became t.54 / t.64
• t.51 🙈 Focus follow: j-arm then immediate k failed to disarm (stale URL-param closure); agent-fixed + unit-tested; overlaps Sol P1-5
• t.52 🙈 Streaming past virtualizer compression threshold nudged scale and jittered rows; agent-fixed via hysteresis; no live repro (unit red only) — withdrawn
• t.53 ❌✅✅✅✅ Arrows-only sample stepper; scroll-position restore on return
• t.58 ✅✅ Fresh-visit-to-top (sibling hop, F5 reload)
• t.59 ❌🙈 Sample-tab scroll-restore drift, reclassified out-of-scope (scout main baseline reproduces broken too)
• t.61 ✅✅✅ Event-card tab preservation (inspect + scout + Cursor)
• t.62 🙈 "Very large sample" parity retest, but Fable wasn't able to supply repro steps
• t.63 ❌✅✅ Focus-page lane breadcrumb dropdown attached to LAST crumb but listed ROOT's sibling lanes; fixed and merged 2026-07-20 ("yes, it is green"); no test covers crumb anchoring.
• t.64 ✅ Finish-scroll fires only on successful runs.
• t.65 ❌✅ c.08 outline bold race: entry stays bold while within its turn scope, across MANY j/k repetitions (the race needs repetition to hit). First pass clean 2026-07-22 ("not spotting a race condition yet") on sample 2025-11-29T13-32-26-05-00_ascii-art-python…3MHJDJPynNSTAVrVL5HPkF.eval sample 1/1, via a ?event= deep link. Peter-verified 2026-07-22: all three round-3 fixes re-verified green, clicked through across both the original audit log and the synthetic clienttl copy, closing the sufficient-repetitions reservation.
• t.66 ❌✅ c.09 agent buttons gone from breadcrumbs and focus view; h/l lane hops still work (Easter egg). Peter-verified 2026-07-22: removal confirmed. PLUS ruling: the h/l regression tests added alongside the removal were "useless slop" — deleted in the slop-sweep rewrite; commit 652bd671 is now production-only. Re-confirmed by Peter 2026-07-22 after fresh pnpm install and the slop-sweep re-SHAs.
• t.67 ❌✅ c.10 switcher renders as pill matching full transcript, inspect + scout. Peter-verified 2026-07-22: "looks good, code diff too." Re-confirmed by Peter 2026-07-22 after fresh pnpm install and the slop-sweep re-SHAs.
• t.68 ❌✅ focus/lane utility-toggle parity in browser: toggle utility ON, focus a utility lane (expect its turn, not an empty page), focus dropdown + h/l lane parity with transcript. A timelines-stripped COPY of the audit log was added at logs/2026-05-06T20-37-46-00-00_audit-clienttl_clienttlcopy1.eval (original left byte-identical, sha256-verified; the copy is synthetic test data — all verdicts below count as "verified on synthetic data only"). Repro steps:
  1. Open http://localhost:5173/#/logs/2026-05-06T20-37-46-00-00_audit-clienttl_clienttlcopy1.eval/samples/sample/ai_consciousness_sycophancy/1 — header must show "14 utility agents hidden".
  2. Click it → utility lanes appear in the swimlanes.
  3. Click a utility lane (or h/l onto it) → scoped view shows "turn 1/1".
  4. Press f → expect a rendered focus page (breadcrumb + model call + turn 1/1), NOT an empty page.
  5. In focus view: breadcrumb dropdown lists the utility lane and h/l steps the same lane set as the transcript.
  6. Toggle utility OFF → behavior identical to pre-change (regression leg).
  7. Copy the utility event's ?event= URL, toggle OFF, open in a fresh tab → the link now force-enables the toggle for that tab and lands on the event (le.48, verified 2026-07-22); a separate fresh tab with no deep link still shows utility hidden.

  Final verdict 2026-07-22 (Peter, clicked through both the original audit log and the synthetic clienttl copy): toggle-parity fix verified — mirror behavior confirmed, utility wrapper turns are reachable in focus view. Testing also confirmed same-named sibling utility wrappers are NOT iterable between each other from within focus view; this is the known lane-model follow-up (le.47), not a defect of this PR. Also Peter's observation recorded: scoring attaches once at sample end, not per utility box — treated as expected behavior unless Peter re-flags it. Standing qualifier holds: all of the above is verified on synthetic data only (see Repro steps above).

• t.69 ⌛ chunked sample renders via ChunkedTranscriptPanel (windowed reads). Repro steps:
  1. Start pnpm dev; open http://localhost:5173/#/logs/2026-07-23T00-00-00-00-00_turnnav-chunked-SYNTHETIC_ChunkedTurnNavA.eval (synthetic fixture, built via the sanctioned chunked converter from a copy of a June turnnav log; 7 event chunks; no existing file modified).
  2. Transcript tab: confirm the chunked marker — a DOM node with data-testid ending in -chunked, or "[chunked] fetch" console lines.
  3. DevTools Network: /api/log-bytes/...?start=&end= byte-range requests, scrolling materializes further events/N.json chunks.
  4. EXPECTED LIMITS, not bugs: no "turn N/N" counters, j/k+Ctrl+G+f inert (le.51), utility indicator absent (le.50).
• t.70 ✅ Fable #13 / commit a8b82eb6: go-to-turn bar jump into a turn hidden by a user collapse. Stack: inspect FE 5173, branch tip 55724231. Repro steps:
  1. Open http://localhost:5173/#/logs/2026-05-06T20-37-46-00-00_audit-clienttl_clienttlcopy1.eval/samples/sample/conspiracy_theory_sympathy/1 — a sample known to render with turn groups.
  2. Find a multi-turn agent-span group in the transcript and collapse it by clicking its event-panel header chevron (the ▼/► icon at the left of the group's header row toggles collapsed/expanded).
  3. Press Ctrl+G (or Cmd+G) to open the go-to-turn bar.
  4. Type a turn number that falls INSIDE the just-collapsed range, then Enter.
  5. EXPECT: the collapsed region auto-expands and the view lands on the REQUESTED turn, with the jump ring marking that turn (pre-fix: it silently landed on the last visible turn before the collapsed region, with no expansion and no ring on the requested turn).
  6. Negative check: repeat steps 3–5 with a turn number OUTSIDE any collapsed range — behavior is unchanged from before this fix (direct jump, no expansion needed).
  Peter-verified 2026-07-23: Ctrl+G into a collapsed range auto-expands and lands correctly.
• t.71 ✅ Fable #10 / commit 7d82d1ce: scanner-result page (ResultBody) no longer flickers chrome while find is open — the branch's one deliberate consolidation change: ResultBody gained the find-open chrome suppression the transcript page already had. Repro steps:
  1. Open any scan result on scout 5174 (route /scan/…, RESULT tab).
  2. Press Ctrl+F to open the find band — it's app-level, keyboard-summoned via useFindBandShortcut, no visible button.
  3. With the band open, scroll inside the results area → EXPECT: the page header does NOT hide.
  4. Close find (Esc in its input, or the close button) → EXPECT: scrolling now hides the header again (baseline behavior, already verified by Peter 2026-07-23).

  Peter-verified 2026-07-23 via A/B against the unfixed bundled viewer (7576): ResultBody behavior is at parity with main with and without find open (header follows scroll position) — ruled good enough. The consolidation kept behavior identical, as intended; the find-suppression added alongside it is not observable live on this surface (see le.37).

• t.72 ❌✅ chunked event-card header pinned to card BOTTOM when scrolled (Peter's screenshot, non-summary tabs). Root cause: the chunked panel positions virtual rows with transform: translateY, and a transformed ancestor traps position: sticky — our branch made card headers sticky-always, but only the monolith list uses absolute-top rows. Fixed (one line, transform → top), commit aac3afbc; browser-verified (jsdom cannot test layout — stated in the commit). Retest steps:
  1. Open http://localhost:5173/#/logs/2026-07-23T00-00-00-00-00_turnnav-chunked-SYNTHETIC_ChunkedTurnNavA.eval (hard-reload — dev server hot-reloaded the fix).
  2. Expand a MODEL CALL card, switch to the API tab.
  3. Scroll within the tall card → the header strip must stay pinned at the top of the scroller, never sink to the card bottom.

• le.01 Port backgroundUpdate protections (branch cross-log-loading-fixes) — lint/typecheck/13 unit tests green, e2e spec written but not executed; sit on new PR until Eric's collection-data rework lands.
• le.02 Focus-link caveats (branch-only events; timeline config mismatch).
• le.03 Scrollbar-drag chrome-ownership gap (pre-existing; Peter ruling pending).
• le.12 Scout backend pin: scout consumes inspect_ai as a PyPI dependency (not a submodule). "Doubly incompatible" = (1) older scout code crashed on an inspect_ai symbol rename, fixed on scout main; (2) scout main's version floor exceeds the newest published inspect_ai, so testing against a dev inspect_ai requires a local editable install (uv pip install -e ../inspect_ai or [tool.uv.sources]). Durable closure: an upstream inspect_ai PyPI release ≥0.3.250, after which scout's scheduled auto-repin bumps the floor and regenerates types. Both failure modes empirically verified by booting 2026-07-23. Local testing unblocked 2026-07-23: scout on main (3fadd555) + venv wired to local editable inspect_ai 0.3.250.dev; FE↔BE handshake proven (branch scout FE loaded 11,338 transcripts via live BE, 3 clean runs, types compatible). Durable closure still = upstream inspect_ai PyPI release then scout auto-repin.
• le.13 Scout hosted-mode checks.
• le.14 Tab-picker label truncation polish — narrow-viewport wrap/overlap fixed 2026-07-20 (title ellipsis, picker floor width, new e2e); further polish tracked here.
• le.18–le.27 Focus-view gaps: load-failure spinner, unanchored pre-turn events, no event-type filter, no sample-level header content, no minimap markers, no streaming indicator, cross-sample paste frame, zero-turn j/k swallow, lane-picker a11y, focus-tab round-trip on exit.
• le.28 Shadow-DOM editable guard (VS Code inputs; overlaps Fable #7).
• le.29 Go-to-turn selection ring (visual retest).
• le.30 Expanded-card gutter (visual).
• le.31 VS Code focus-ring suppression.
• le.32 Scout turn-nav writes ?event= (smoke check).
• le.33 Summary-header expand asymmetry (inspect position-based, scout direction-based) — kept as-is, intentionality unresolved.
• le.34 ✅ e2e suite fake-green audit — Playwright output piped through tail swallowed failures and exit codes. RESOLVED 2026-07-22: the full e2e suites ran with real exit codes captured directly from the runner process (no truncating pipes): apps/inspect 95/95 passed exit 0, apps/scout 64/64 passed exit 0 (159 total), on dedicated ports without touching the live dev servers, MSW-mocked. The previously-scheduled run happened; audit closed.
  Adjacent: a full-workspace turbo run 2026-07-22 on the branch post-install came back 26/27 tasks green; the sole failure was the le.49 flake (not a PR-354 regression).
• le.36 Back-from-focus lands at top when entered from passive scroll (c.02 divergence).
• le.37 Scout find-band ownership in focus/scanner views. 2026-07-23: the find-open chrome suppression added to ResultBody by the round-4 consolidation is not observable live (position-based expand-only-at-top wins); parity with main ruled good enough — revisit only if scanner-view find UX gets real attention.
• le.38 pointerdown cancels click-initiated settles (needs design).
• le.39 Scout visit-keyed state bags never evicted.
• le.40 Turn-count derivations: 3 call paths (outline, event-card counter, focus view) use different collapse/scorer params; behavior consistent under 2026-07-20 testing; candidate post-PR refactor to one parameterized derivation.
• le.41 h/l lane hops land on the lane's turn 1 instead of transcript-top with collapsed headers (?event=-router side effect); design revisit post-merge.
• le.42 ?message= deep-link settle — replaced main's mechanism with a simpler one; both fail the same way on very deep messages, pre-existing on main. Overlaps Fable #4.
• le.43 kAtTurnTopTolerancePx = 24 — uncalibrated constant, accepted as-is ("maybe it's fine").
• le.44 Upstream candidates: t.59 debounce fix; inspect_ai LFS resolver mixed-dir assert; scout repin; inspect_ai's Python timeline writer gaining utility-agent classification parity with ts-mono, so logs it writes carry the "hidden" flags without depending on the client-side classifier (ruled 2026-07-22 as out of scope for this PR); scout: an orphaned swimlane selection key persists in the URL hash (?selected=<key>, unlike inspect where the equivalent selection is transient and cleared on click), making it a shareable/bookmarkable landing state that can hit the raw-stream fallback below — strengthening the case for fixing that fallback upstream. Second instance 2026-07-23: scout view's dist LFS resolver asserts all-files-are-pointers; stray built .js.map files in scout's ts-mono dist break boot until removed (rebuilding scout's bundled viewer regenerates them).
• le.45 Hidden-turns items deferred out of this PR: the pre-existing (branch-worsened) bug where toggling utility off leaves an orphaned selection falling back to the raw event stream, inflating outline/turn counts until the selection is clicked away (upstream candidate, overlaps le.44); a narrow edge case where the utility toggle is the only thing that changes a turn count (inline-agent-id shape); and a follow-up check that scout's handling of server-baked timelines has the same classification parity as inspect's.
• le.46 Outline bold near the log TAIL: the two selection writers are now both valid but can pick DIFFERENT rows near the end (scroll tracker keeps advanceDetectionPointAtEnd=true, j/k tracker sets false; last-writer-wins). Pre-existing; also pre-existing kFramesToStabilize=10 rAF settle latch in useOutlineScrollSync deliberately untouched (uncalibrated timing guard, separate change if ever).
• le.47 Lane-model redefinition, follow-up PR after #354 merges, decided 2026-07-22: one rule — "a focus/h-l stop is every swimlane box." Per-instance lanes for same-named sibling agent spans (a row that currently aggregates e.g. 9 utility-wrapper instances into one stop becomes 9 stops; focus header shows e.g. "utility k/9"); phase rows (scoring) admitted as their own stop — scoring is currently appended to a turn's slice instead, dropped from the focus terminal-append once phase rows exist (error/sample_limit stay appended); every lane anchored on its first event (model or not). A scan of all 1,547 eval files / 12,023 samples in Peter's logs found ZERO non-utility rows with more than one same-named sibling agent span (and zero server-timeline samples) — the sibling-instance gap is utility-only in practice today. A net-code-deletion hypothesis for this change (fewer special cases elsewhere) was tested and REFUTED: only ~5–10 lines are deletable, additions make the diff neutral-to-positive overall; the payoff is conceptual uniformity, not code shrinkage. Sized ~120–180 LOC across 4–5 files redefining the lane contract both views consume. Keying landmine noted: FocusLane.laneKey currently shared across siblings; the :N sub-selection infra already exists in timelineEventNodes.ts. Non-model scorer lanes need new anchoring logic.
• le.48 A shared deep link (?event=) to a utility-wrapped event, opened by a recipient whose utility toggle is off, still resolves to nothing silently — on both surfaces that consume ?event=: the transcript (useDeepLinkResolution) and the focus event page (useFocusLaneScope; caught during browser retest, transcript-only in the first cut). Ruled 2026-07-22: such a deep link should force-enable the utility toggle for that visit rather than resolve to the enclosing lane or show an error — links must never silently no-op, matching the precedent set by ?follow=1 overriding local state. Implemented in-branch, folded into the 4th commit 1b70adbb, covering both surfaces. Red-first tested on both (transcript: setIncludeUtility never called pre-fix, verbatim Number of calls: 0; focus page: verbatim expected false to be true // Object.is equality). All four gates green. ✅ Peter verified both surfaces 2026-07-22: the previously-dead focus-page and transcript ?event= links now flip the toggle and land correctly, and the flip stays scoped to the visiting tab (a fresh tab still shows utility hidden — no cross-tab persistence leakage, matching the ruling's "for that visit" intent).
• le.49 @tsmono/react test "lazy mathjax loading > retries the mathjax import after a failed chunk load" times out (hardcoded 5000ms) under full-workspace parallel test load; passes standalone (819ms) and in package isolation (173/173). Pre-existing timing fragility, not PR-354-related; candidate fix: raise/remove the timeout or isolate the test.
• le.50 Chunked samples (#451, internal-only) run the client classifier over synthetic empty-payload events — utility detection never fires, indicator shows 0, toggle no-ops; fidelity gap between #451 and #425, not caused by this PR; upstream/product question.
• le.51 Turn navigation is entirely absent on chunked samples — ChunkedTranscriptPanel never mounts the keyboard hook, go-to-turn bar, or focus navigation (verified red/green against a monolith control of the same eval). Investigated 2026-07-23 with a red probe against the real pipeline: the chunked skeleton assigns all model calls within one gap the SAME ordinal (the fixture's 100 model turns collapse onto 4 distinct ordinals), so j/k / Ctrl+G via ordinals would land wrong for most turns — worse than no navigation at all. A safe implementation needs a new row-scan navigation design over materialized rows, and Ctrl+G-by-turn-number is impossible from skeleton data alone. DECLINED for this PR on that basis; follow-up only, with this ordinal-degeneracy constraint as the starting point. Needs raising with the reviewer alongside le.50 as one "chunked samples" caveat.
• le.52 An unpushed 2026-07-17 agent commit in inspect_scout (fix/transcripts-discovery-mkdir, 1ca4e6c7) is an UNVERIFIED WORKAROUND, not a vetted fix: part 1 (no mkdir on read-only transcript discovery + OSError handling) is a defensible idea needing independent review — scout main still has that bug but it only bites on inaccessible paths; part 2 (skip per-request coverage re-scan) is a latency hack suppressing staleness warnings — drop unless justified. Review before any upstream use; nothing cherry-picked 2026-07-23.