perf(codex): stream rollout metadata instead of parsing every line

[od0]

Found while dogfooding bare path resume (#154): the pre-picker gather stalled for 80 seconds with no output. /usr/bin/time pointed at codex — every session-listing surface (p list codex, the share picker, bare resume) calls read_metadata per rollout file, which full-parsed every JSONL line through serde. On a real tree (421 rollouts, 3.9 GB, 89 files >10 MB) that's ~80 s of pure parse CPU. The old code even predicted this: "If that becomes a bottleneck we'd peek the first line plus stat for mtime."

The fix

read_metadata is now a single streaming pass that exploits the rollout being an append-only log:

• Head (bounded, ≤100 non-empty lines parsed): session_meta (id/cwd/cli_version/git), first timestamps, first user prompt — all live at the top.
• Tail (1 line): the newest timestamp.
• Everything between: counted, never JSON-parsed.

Measured on the tree above: p list codex 1m20s → 8.5 s in a debug build (release is faster); user-time drops from 72 s to ~6 s, i.e. it's raw I/O now.

Deliberate trades (documented on the method)

• first_user_message is None if the first prompt appears after the head budget (real sessions surface it within a dozen lines).
• line_count counts non-empty lines (unparseable ones included) instead of successfully parsed ones — it only ever approximated message_count.
• started_at/last_activity come from head-min/tail-max rather than a global min/max scan — identical for append-only logs.

Three new tests pin the tail-timestamp path, tolerant counting over blank/junk lines, and the head-budget bound. Also picks up the pre-existing paths.rs rustfmt drift that fails cargo fmt --check on main under the pinned 1.94.0 toolchain.

toolpath-codex 0.6.1 → 0.6.2 (perf fix, patch bump) in the four conventional places.

Follow-up worth filing separately: the listing surfaces re-scan all files on every invocation; piggybacking session metadata on the sync manifest's stat stamps would make an unchanged tree instant.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

🔍 Preview deployed: https://009fd271.toolpath.pages.dev

[od0]

Folded into #157 — the codex metadata streaming commits are on that branch verbatim (759ced0), part of the single picker-performance story with the parallel gather and the listing cache (#158). No review needed here.

[od0]

Reopened as the bottom of stack #160 (#156 → #157 → #159 → #154) — same commits as before the fold, now reviewable layer by layer. gh stack merge from the top handles the chain.

[akesling]

This should be pulled out to a helper for readability.

[akesling]

This code looks rather fragile.

What performance do we get if we use serde over a partial struct with raw values / accepting "undefined" fields? If serde won't give us what we want out of the box, we may want to roll our own partial Visitor or a proper parser that skips things intentionally.