fix(turn-budget): one turn, one number — hook, bell and panel parity

[axisrow]

Problem

Session 67105434 showed Turn budget · 15.1M / 15.0M in the notification bell, while the Visible Context panel showed Re-read · Turn 13 · 60 rq · 7 425 348 tok. Half the spend seemed to vanish between the two surfaces, and there was no way to tell which turn each number belonged to.

Live-file audit (scripts/turn-accounting canon): both numbers were correct — for different turns. 15.1M was the running total at the moment turn 3/8 crossed the 15M budget (final sums 15.46M / 15.63M, both in earlier context phases, so the panel — which shows the current phase — couldn't show them). 7.43M is the panel's Turn 13, exact to the token. The notification simply never named its turn.

Changes

• Main-chain filter in the canon (scripts/turn-accounting.mjs): new isMainChainAssistantLine (same semantics as loopDetection.ts:314); analyzeTurn applies it, the CLI walker (turnSpendStats.ts feedLine) imports it — the hook, the detector and the panel now exclude sidechain/synthetic lines by the same definition.
• Notification names its turn (loopDetection.ts, FileWatcher.ts): TurnBudgetIncident carries turnNumber + turnStartTs; the message reads Turn budget · turn 8 (07:08) · 15.1M / 15.0M — … (falls back to current turn when the file opened mid-turn).
• Panel matches (contextTracker.ts, RereadSection.tsx): RereadInjection.turnStartTs → each Re-read row shows the turn's start time next to Turn N, so a notification and a panel row can be matched unambiguously.
• T2 — parity pinned on real-shaped data: turnAccounting.parity.test.ts now asserts analyzeTurn (hook) == TurnBudgetDetector.feed fed in batches of 1/3/5 at every prefix (the live FileWatcher shape) == sumTurnReread (panel), plus sidechain exclusion everywhere. Detector got a currentTurnTotal seam for it.
• T3 — --audit CLI: pnpm turn-spend:stats -- --audit <file.jsonl> prints the three accountings side by side per turn and a parity verdict. Live runs: parity OK on 19 + 278 + 82 turns across three sessions, including the 15.46M / 15.63M turns.

Number inventory (what is covered where)

Number	Shown at	Computed by	Parity
turn re-read	hook deny, bell notification, panel Re-read, chat pill	turn-accounting.mjs (single canon)	parity test + --audit (this PR)
turn budget spent-at-crossing	bell notification	TurnBudgetDetector (same canon, edge snapshot)	covered by the incident test
per-turn billed chips, session totalTokens/turnCount	chat chips, sidebar, CLI	separate consumers of the canon	not covered yet — candidates for follow-up issues

Fixes the "where did 15M go" confusion; the numbers themselves were never wrong.

[axisrow]

Verdict: changes requested

The accounting core is solid. Verified: the canon filter matches the detector predicate; the --audit indexing is correct (detTotals[0] is the pre-first-boundary prefix, one push per boundary plus the final push, so detTotals[i+1] is the final total of turn i+1); the batch-feed parity test exercises real prefixes; the ChatHistory merge of rereadAll cannot double-count (reread is excluded from totalEstimatedTokens, the pill total comes from lastAssistantTotalTokens). The jsonl.test.ts fixture change is correct and required: after #67 mixed relay+text is user text (the whole-content rule makes isUserChunkLine true), so both edited tests fail on current main — the pure-relay fixture restores their original intent.

Two findings, both on the core deliverable of this PR — matching a bell notification to a panel row:

1. Turn numbering bases diverge (inline, loopDetection.ts). The detector increments turnNumber on canon isTurnBoundary = user messages + compaction summaries. The panel and chat Turn N number only user messages + teammate relays (aiTurnIndex(userCount + relayCount); jsonl.ts turnCount explicitly gives compaction no number). A compaction before the crossing turn inflates the bell number by the compaction count; teammate relays deflate it. Session 67105434 — the session that motivated this PR — had compaction phases, so the mismatch is guaranteed in exactly the target scenario. The canon already provides isTranscriptTurnLine as the chat Turn N semantics; number the incident with it while keeping the spend-bucket reset on isTurnBoundary.

2. Row time not locale-pinned (inline, RereadSection.tsx). Commit d92e999 pinned the bell time to 24h with hour12:false for a locale-independent message, but the panel row renders without it — in a 12h locale the panel shows 7:08 PM against the bell 07:08. Reuse the same options.

Notes, no change required: the --audit comment claims the detector path is exactly what FileWatcher streams, but it feeds raw JSONL lines rather than ParsedMessage — accounting-equivalent, wording only. The new parity tests cannot catch finding 1: the fixture has one turn and no compaction or relay, where both numbering bases coincide. Consider a preview-layer test that mixed relay+text IS picked as firstUserMessage — the #59 canon is currently untested at that layer after this fixture swap.

[axisrow]

Verdict: approved

Re-review of 7feaa7b. Both findings from the previous round are addressed exactly along the suggested direction, and the new tests pin the semantics.

1. Turn numbering (loopDetection.ts): the detector now numbers turns with isTranscriptTurnLine minus isSidechain/isCompactSummary guards — the jsonl turnCount / chat chip canon — while the spend bucket still resets on isTurnBoundary, carrying the number and start timestamp across the reset. Verified the ordering: numbering increments before the bucket swap so the fresh bucket carries the incremented number; a relay advances the number without resetting the bucket; a compaction resets spend without advancing the number (the guards are needed because isUserChunkLine returns true for masquerading compact summaries — the code comment documents this). Compaction no longer inflates the bell number, relays no longer deflate it; the number now matches the panel Turn N and chat chips in the compaction and team-relay scenarios that motivated the PR.

2. Row time (RereadSection.tsx): hour12:false added — the panel row and the bell now render identical locale-independent times.

New tests cover both semantics with correct arithmetic: the relay test pins number 2 with spend flowing across it (6M + 6M = 12M in one bucket), the compaction test pins number 1 with the pre-compact 9.5M dropped and the crossing at 10.5M post-compact. Shared re-export of isTranscriptTurnLine verified.

Residual notes from the previous round remain optional and non-blocking: the --audit wording about the detector path, and a preview-layer test that mixed relay+text is picked as firstUserMessage. No changes required.