Closes #171

[realproject7]

Closes #171

Reset every session-scoped webview singleton when a new session starts, so a second session in the same app run no longer inherits the first's feed, cards, summary/board, or archive path.

EPIC Alignment

Part of EPIC #134 (v1.3 audit follow-up), Batch 39, standing-order sequence position #171. Webview live-view cluster; scoped to src/main.ts + two small DOM/Tauri-free reset seams: FeedState.reset() (src/feed-state.ts) and a new SessionScope (src/session-scope.ts) grouping the session-scoped summary/board/archive-path/metrics values. RE1's ticket-review (TICKET-OK, chat #876) confirmed the reset boundary is the session://status phase transition and the scope is bounded to src/main.ts. Does not touch resample.rs/render.ts/parse.ts/pipeline.rs (reserved for #178/#179) or any dashboard/settings files.

Self-Verification

The bug. The webview's session-scoped state is created once at module load and never cleared. The supported flow is Stop → idle Start screen → Start again in the same running webview (no reload). So session 2 started with session 1's state:

• session 1's caption blocks were still in feed/DOM and reappeared once phase left idle, with new captions appended after them and no boundary;
• openReview() (main.ts:856-857) builds its coaching list from feed.micUtterances(), which still held session 1's own utterances — so session 2's post-meeting review/coaching could rewrite lines actually spoken in the previous meeting;
• latestArchivePath/latestSummary/latestBoard could show session 1's values in session 2's review.

The fix — two unit-tested reset seams + a DOM reset, run at the new-session boundary.

• FeedState.reset() (feed-state.ts:54-60) clears blocks (in place, so live shimmer/heartbeat references see the empty feed), byId, partials, keyCounter, and the eviction count.
• SessionScope (session-scope.ts:26-41) groups the five session-scoped model values that live outside the feed — summaryLine, latestSummary, latestBoard, latestArchivePath, pendingMetrics — behind one reset(). main.ts holds one scope (main.ts:90) and reads/writes scope.<field> (summary strip, openReview, the host-event handlers).
• resetSessionState() (main.ts:1087-1105) runs both seams (feed.reset(), scope.reset() at :1104) plus the DOM/rAF reset: remove blockEls nodes (the permanent #feed-note stays), feedCoalescer.drain(), shimmerEl/atBottom/feedNote, the pendingCards/analysisCards result cards (cardsEl.replaceChildren() + clear maps), then renderPinned().
• It fires at the transition into a new session (main.ts:1116-1117: a non-active prevPhase → starting|live), guarded so pause/resume and starting→live never re-clear a live session — and crucially not on Stop, so the just-ended session's review screen keeps the data it already captured. render() + syncMiniViews() then repaint the clean panel.

Session-scoped state audit (Scope requirement — every module-level singleton in src/main.ts):

State	Reset?	Where
feed (FeedState)	✅ feed.reset()	blocks/ids/partials/keys/eviction — unit-tested
blockEls (Map + DOM)	✅ remove nodes + clear()	keeps #feed-note sentinel
feedCoalescer	✅ drain()	discards queued block writes
shimmerEl	✅ → null	old node removed with blockEls
atBottom	✅ → true	scroll-stick to live edge
feedNote visibility	✅ removed	history notice hidden
pendingCards (Map + DOM)	✅ cardsEl.replaceChildren() + clear()	reply/quick-translate cards
analysisCards (Map + DOM)	✅ same	#80 analysis cards
summaryLine	✅ scope.reset()	SessionScope — unit-tested
latestSummary	✅ scope.reset()	SessionScope — unit-tested
latestBoard	✅ scope.reset()	SessionScope — unit-tested (fresh object, no shared-ref bleed)
latestArchivePath	✅ scope.reset()	SessionScope — unit-tested (no stale saved-file path)
pendingMetrics	✅ scope.reset()	SessionScope — unit-tested
coachingCards	n/a	view into the review surface, dismissed at the same boundary (review.hide())
channels	n/a	re-seeded every start by session://channels
requestCounter	intentionally kept	monotonic → card ids stay unique across sessions
feedFlushScheduled	n/a	self-clearing rAF guard (a pending flush drains the empty feed → no-op)
phase/statusDetail	n/a	driven by the status stream itself

Security invariants. No new dependencies. No caption/audio content logged — resetSessionState()/FeedState.reset()/SessionScope.reset() emit no logs. feed stays const (reset via method, no reference churn); no drive-by refactors. Only src/main.ts, src/feed-state.ts, the new src/session-scope.ts, and the test touched.

Tests / evidence. test/session-reset.test.ts (6 tests) drives the exact start → content → stop → start sequence against both reset seams:

• feed (3): session 1 mic+them finalized captions, a translation, a pin, a live partial → after reset() the feed is clean — no blocks, no leftover mic utterances (the coaching-leak core), no pins, no stale ids, eviction 0, render-key counter reset.
• SessionScope (2): populate summaryLine/latestSummary/latestBoard/latestArchivePath/pendingMetrics (exactly what openReview reads) → reset() → assert all cleared, incl. a no-stale-archive-path assertion and a fresh-board (no shared-reference bleed) case. This is RE1's P1: the review inputs are now provably cleared.
• combined (1): feed + scope reset together, mirroring resetSessionState's model half → fully clean session 2.
• Only the block/card DOM node clearing stays exercised in the running app (main.ts is DOM/Tauri-coupled; no main.test.ts in-repo), routed through the same resetSessionState.
• Full app suite green: 153 tests / 22 files pass (pnpm test:app); pnpm typecheck, pnpm lint, and pnpm build all clean.

Design Fidelity

No new visual surface: the reset produces already-designed empty/idle states, repainted by render()/syncMiniViews() at the session boundary. Each state is spot-checkable against its existing source; nothing in this PR adds or alters CSS, so there are no layout, token, spacing, color, or animation changes.

State after reset	Expected (existing design)	Source pointer	Layout/token/anim change?
Feed (empty)	#feed holds only the permanent #feed-note sentinel; no caption blocks	styles.css:376 (#feed), :383 (#feed-note); reset removes tracked blockEls nodes only (main.ts resetSessionState)	None — same markup, no rule touched
Feed history notice	#feed-note hidden (no evicted history yet)	styles.css:383,389 (.visible toggle); reset clears .visible	None
Pinned dock (empty)	#pinned without .has-pins → collapsed dock	styles.css:570,576; reset → renderPinned() over the empty feed removes .has-pins (main.ts:526-529)	None
Result/analysis cards (none)	#cards empty container	styles.css:585,591 (#cards/.card); reset cardsEl.replaceChildren()	None
Summary strip (cleared)	#summary label/line reset to the idle/starting content, dot off	styles.css:324,331,337; renderSummaryStrip() reads the cleared scope.summaryLine (main.ts:394-399, called by render())	None
Strip/Capsule mini-views	latest line falls back to Listening…/LiveCap (empty feed)	syncMiniViews() reads feed.latest()	None

The only user-visible effect is that session 2 starts clean instead of showing session 1's leftover transcript/cards/summary; every rendered state is one already in the design system.

Deviations

• Extracted SessionScope (src/session-scope.ts) for the five summary/board/archive/metrics values, and FeedState.reset() for the feed — two DOM/Tauri-free seams so the whole model reset is unit-testable (RE1 P1 on the first revision: the review inputs were previously reset only inside DOM-coupled main.ts and thus unproven). feed stays a const (reset via method, zero reference churn); main.ts now references scope.<field> for the grouped values.
• The reset fires on session start, not stop, specifically so the just-ended session's review screen (opened on stopped) still reads its summary/board/utterances/archive path.

🤖 Generated with Claude Code

[project7-interns]

Verdict: REQUEST CHANGES

Epic Alignment: NOT REVIEWED

Structural gate failed before code review.

Checked (evidence)

• Live PR body contains filled ## EPIC Alignment and ## Self-Verification sections.
• UI structural requirement: ## Design Fidelity is prose, not the required table.

Findings

• [P1] Required Design Fidelity table is missing
  • File: PR body, ## Design Fidelity
  • Why it fails: this webview-facing change must provide the required table so the reviewer can verify design/state coverage. Narrative text cannot be spot-checked row-by-row and does not satisfy the template.
  • Do instead: replace the prose with a table covering the affected empty/idle feed, pinned dock, cards, and summary states, each with a source pointer and a statement that no layout/token/animation behavior changes.

Decision

Add the required table and re-request review; no code review has been performed.

[realproject7]

RE2 — APPROVE at PR #182 head (pending app-macos/packages-linux green)

The #171 session-scoped reset is correct, the state audit is complete, and the real leak is fixed. Reviewed the whole diff (src/main.ts, src/feed-state.ts, test/session-reset.test.ts) and independently ran the app suite + typecheck + lint on Linux.

Checked (evidence)

• The leak is real and fixed. openReview (main.ts:859-860) builds its coaching utterance list from feed.micUtterances(); with feed a never-reset webview singleton, session 2's review read session 1's mic utterances, and latestArchivePath could show session 1's path. resetSessionState (main.ts:1091-1113) → feed.reset() (feed-state.ts:54-60) clears feed.blocks (mutated in place so live shimmer/heartbeat refs at main.ts:521 observe the empty feed), so session 2 starts clean.
• Reset fires at the correct edge, exactly once. wasActive = prevPhase ∈ {starting,live,paused}; reset runs when !wasActive && phase ∈ {starting,live} (main.ts:1124-1125). Traced the phase machine: idle→starting fires; starting→live and paused→live (resume) do not re-clear; live→paused/stopping, stopping→idle do not fire; stopping→starting (stop→start) fires. Not on Stop, so the just-ended session's review keeps its captured data. ✓
• State audit independently verified complete. Enumerated every module-level mutable singleton (let + the 3 Maps). Reset: feed, blockEls (+DOM .remove()), pendingCards/analysisCards (+cardsEl.replaceChildren()), summaryLine, latestSummary, latestBoard, latestArchivePath, pendingMetrics, shimmerEl, atBottom, feedNote visibility, feedCoalescer.drain(). Correctly kept: channels (user mic/system config, re-seeded by session://channels :1135), requestCounter (monotonic → unique card ids), feedFlushScheduled (a self-clearing rAF guard :494-500, holds no session content), toastTimer/chromeTimer/dragStart (transient UI), capabilities/appSettings/boot*/onboardingDecided (app-level). No session-content singleton is missed.
• coachingCards (main.ts:832) is a review-owned registry keyed by the monotonic requestCounter, rebuilt per openReview from the (reset) feed and dismissed via review.hide() at the same edge — no stale-content bleed.
• Independently reproduced (Linux, node 24 / vitest 3.2.6): full app suite 22 files / 150 tests pass; new test/session-reset.test.ts 3/3 — incl. the exact leak assertion (session 1 has 2 mic utterances → after reset(), micUtterances() is []), key-counter reset, and live-partial drop. App-tsconfig tsc --noEmit clean; eslint on the 3 changed files clean.
• Boundary/invariants clean: only the 3 files, package.json/lock untouched (no new deps), no resample.rs/render.ts/parse.ts/pipeline.rs or dashboard/settings touched, no caption content logged.

Notes (non-blocking)

1. Test coverage boundary: the FeedState.reset() model half is unit-tested; the main.ts DOM/scalar resets (blockEls/cardsEl/summary/board/archivePath/metrics) are DOM/Tauri-coupled with no main.test.ts in the repo, so I verified them by inspection (each cleared, correct edge). Acceptable given the repo's existing test surface; the model-level leak — the audit's actual finding — is the one that's unit-covered.
2. Reset-before-events ordering relies on the host emitting session://status starting/live before session-2 caption events. That's the normal host lifecycle; JS runs the status handler (incl. reset) to completion before the next event.

CI condition

color-guard/no-stub-gate pass; app-macos/packages-linux pending — packages-linux runs the authoritative full lint/typecheck/test:app (incl. the host tsconfig). My Linux app-suite repro de-risks it; the merge gate should still wait for both green. (Shared bot token can't file a formal GH approval; this comment + chat is my evidence-bound verdict.)

[project7-interns]

Verdict: REQUEST CHANGES

Epic Alignment: PASS

The one-file webview-state reset plus FeedState model reset advances #171 without touching the reserved #178/#179 files.

Checked (evidence)

• Structural gate: live PR body now has filled alignment/self-verification sections and a six-row Design Fidelity table.
• Design table spot-check: the cited reset paths match src/main.ts:1091-1113; no CSS/layout/token/animation changes appear in the live diff.
• Reset behavior: src/main.ts:1091-1129 clears the model, DOM/maps, coalescer, and review inputs only on a non-active → starting/live edge.
• Feed model: src/feed-state.ts:54-67 clears blocks, ids, partials, render keys, and eviction count in place.
• Riskiest part of this diff: stale session data crossing the start boundary; the runtime reset covers it, but the required scalar/review-state regression proof is absent.
• Kill-list: scanned new ranges — clean.
• CI: gh pr checks 182 → app-macos, color-guard, no-stub-gate, packages-linux all pass.

Findings

• [P1] Regression coverage omits required summary, board, and archive-path reset
  • File: test/session-reset.test.ts:29-91
  • Why it fails: all three tests call only FeedState.reset(). They never execute resetSessionState() or assert summaryLine, latestSummary, latestBoard, or latestArchivePath; therefore they do not cover #171 AC’s start → feed content → stop → start with clean summary/board and no stale archive path. The PR body explicitly labels those runtime resets as inspection-only.
  • Do instead: add a testable reset-state seam (or DOM/Tauri test harness) that drives the status transition and asserts those four review/UI values are cleared alongside the feed.

Decision

The runtime code and fidelity documentation are aligned, but the stated regression acceptance criterion is not yet tested end-to-end enough to approve.

[project7-interns]

Verdict: APPROVE

Epic Alignment: PASS

The #171 fix resets session-bound webview data at the new-session edge and now test-covers both the feed and review-state models without expanding into reserved files.

Checked (evidence)

• Re-review delta: 8de4d3a421d820184b8e69b279103e91e3055d31..2226e1d7f066a897d973800713d51f68ad64e4df.
• Reset wiring: src/main.ts:1080-1104 calls scope.reset() with the existing feed/DOM cleanup at the guarded new-session boundary.
• Review-state model: src/session-scope.ts:26-41 resets summary line, summary, a fresh board, archive path, and metrics.
• Regression coverage: test/session-reset.test.ts:94-160 proves those values and the feed clear together on the modeled start → content → stop → start flow.
• Design Fidelity: live PR table spot-checked against src/main.ts:1091-1113; no CSS, layout, token, or animation changes in the diff.
• Riskiest part of this diff: stale review inputs surviving between sessions; the new SessionScope seam is exercised directly and wired to the runtime reset.
• Kill-list: scanned new ranges — clean.
• CI: gh pr checks 182 → app-macos, color-guard, no-stub-gate, packages-linux all pass.

Findings

• None.

Decision

The prior test-coverage gap is resolved, and the scoped reset behavior is supported by live green CI.

[realproject7]

RE2 — APPROVE at 2226e1d (delta re-review; app-macos now green per #931)

@re1's P1 is genuinely resolved: the summary/board/archive/metrics reset is now a DOM/Tauri-free unit-tested seam, and the SessionScope extraction is a complete, behavior-preserving refactor. Re-reviewed the delta only (8de4d3a..2226e1d) and independently re-ran the suite + typecheck + lint on Linux.

Checked (evidence)

• P1 closed. The five scalars moved into SessionScope (session-scope.ts:26-41) with a reset(); resetSessionState now calls scope.reset() (main.ts:1104) alongside feed.reset(). New tests prove the clears: SessionScope.reset populates all five (the exact set openReview reads) → reset() → asserts all cleared, including an explicit no-stale-archive-path assertion; a fresh-board not.toBe(board1) case; and a combined feed+scope start→content→stop→start test mirroring resetSessionState's model half.
• Refactor is complete and behavior-preserving. All 17 scope.<field> references migrated (summary strip :391, openReview :860-869, host-event handlers summary/metrics/archived/stopped :985-1055); zero orphaned bare references to summaryLine/latestSummary/latestBoard/latestArchivePath/pendingMetrics remain (grep-verified), and the now-unused BoardWire import is correctly dropped. feed stays const (no reference churn). emptyBoard() returns a fresh object each call — no shared-reference bleed (unit-covered).
• Independently reproduced (node 24 / vitest 3.2.6): test/session-reset.test.ts 6/6; full app suite 22 files / 153 tests pass; app-tsconfig tsc --noEmit clean (covers main.ts + session-scope.ts); eslint on the 3 changed files clean.
• Boundary/invariants still clean: src/main.ts + src/feed-state.ts + new src/session-scope.ts + the test; package.json/lock untouched (no new deps); no reserved files; no caption content logged.
• Everything from my prior full review still holds (leak fixed, reset fires on the correct once-per-session edge, full state audit complete) — this delta only refactors the five scalars behind a seam and adds tests; it does not change the reset semantics.

Residual note (non-blocking, unchanged)

Only the block/card DOM-node clearing in resetSessionState stays inspection-verified (DOM/Tauri-coupled, no main.test.ts in-repo) — but both model halves (feed and scope) are now fully unit-tested, which was exactly @re1's ask.

CI

app-macos/color-guard/no-stub-gate/packages-linux all green at 2226e1d per #931 — my CI condition is met. (Shared bot token can't file a formal GH approval; this comment + chat is my evidence-bound verdict.)