fix(codex): exclude subagent fork replay history from token accounting (#75)

[pitimon]

What & why

Closes the Codex subagent fork-history overcount (#75). A spawned Codex subagent rollout (session_meta.payload.source.subagent.thread_spawn) replays the parent thread's cumulative token_count history before a deterministic inter_agent_communication_metadata boundary, then the genuine child turns. The parser counted every replayed parent delta as the child's consumption — inflating Cache Read by up to ~120x on multi-agent workflows (the replay can be ~99% of the file).

Root cause (verified against real local rollouts)

The cumulative total_token_usage counter is continuous across the boundary — the last replayed cumulative and the first child cumulative differ only by the child's genuine first turn. Example from a real fork file: last replay total=701736 → first child total=726896 (Δ 25160 = the child's real first turn). So the correct fix is baseline-tracking, not a reset.

The fix

• In replay mode: advance the delta baseline (totals = totalUsage) but attribute nothing.
• At the boundary: flip accounting on. The first child delta = firstChildCumulative − lastReplayCumulative — robust even if last_token_usage is absent.
• src/lib/rollout.js (parseRolloutFile, the incremental path behind dashboard Cache Read): the fork lifecycle (null/replay/child) is persisted in the per-file cursor as codexFork and threaded through parseRolloutIncremental. A partial fork with no boundary yet fails closed — cursor held at its start offset, incoming baseline returned unchanged (never poisoned) — so the completed file is re-read next pass.
• src/lib/codex-rollout-parser.js (parseCodexRolloutFile, context-breakdown drill-down): same baseline-tracking; replay also skips collecting pending tool/exec calls. Fork-control records are handled before the date/timestamp gate so a query window can never drop the boundary and leak replay back in.
• The non-fork path keeps its original guard order (the replay-only advance is scoped to forks), so no other provider's accounting changes — locked by a regression test against a cumulative dip.
• Stuck/partial forks are surfaced under TOKENTRACKER_DEBUG; the record-type string literals are marked as a Codex format contract.

Local-only guarantee

Parsing and tests read only local files; nothing is sent anywhere (per #75's privacy requirement).

Empirical result (measured on the primary path)

Ran the fixed parseRolloutIncremental over 153 real fork rollouts in ~/.codex with an empty cursor, versus the pre-fix main parser over the same files:

	Cache Read (cached_input)	Total
Before (main)	8,497,618,432	8,734,717,609
After (this PR)	204,393,984	217,757,279

~41.6x Cache Read inflation removed across the local fork corpus (the issue's 120x was one 6-file workflow; this is the aggregate over all local forks). The drill-down parser drops the same file's total_tokens from 1,922,762 → 1,221,026 (the ~700K replay excluded).

Note: this PR fixes go-forward parsing. Already-ingested aggregates keep their inflated values (those files sit at offset = EOF and are not re-read) until the follow-up migration recomputes them. #75 stays open after this merges.

Test plan

test/codex-fork-history.test.js — 9 regression tests (issue's 5 fixtures + extras):

• normal (non-fork) rollout unaffected
• non-fork cumulative dip does not re-baseline (no over-count on recovery)
• complete fork counts only post-boundary child turns
• partial fork (no boundary) ingests nothing and does not advance the cursor
• fork resumes correctly once the boundary is written
• child turns appended in a later incremental read counted exactly once
• multiple subagents forked from the same parent are isolated
• drill-down parser child-only (plain + day-scoped window)
• npm run ci:local green (root suite 760/760, validators, dashboard build)

Follow-up (separate PR, per review)

Idempotent one-time rebuild of already-inflated Codex queue + project aggregates (reset codex cursors + recompute) with a no-network acceptance test. Staged separately because it mutates persisted local data.

Refs #75