pi/omp: discover transcripts whose session record follows a title slot

[avs-io]

This overlaps #846, which came first. @jbspeakr — who reported #845 — opened #846 fixing the same defect. This branch was built independently from the issue before #846 surfaced in a PR sweep. #846 should have precedence.

Left open as a draft only for the one difference described below. If that difference is not wanted, close this in favour of #846.

Relates to #845.

The defect

The shared Pi/OMP JSONL discovery gate inspected only the first physical line of a transcript for a type: "session" record. Oh My Pi writes a fixed-width type: "title" metadata line ahead of that header, so valid OMP transcripts were excluded from codeburn sessions --provider omp. Both this branch and #846 scan past leading non-session lines to fix it, and both keep Pi and OMP on the identical shared path.

Difference from #846

Read cost. #846 keeps readSessionFile, which reads the entire transcript into memory (up to the 128MB cap) and then iterates its lines. This branch streams through readSessionLines — the helper codex.ts, droid.ts and others already use — and stops after 20 leading lines.

The distinction matters most for the negative case. A message-only file with no session record is read in full under both the old code and #846; under a bounded scan it costs 20 lines. Session directories accumulate such files, and discovery walks all of them on every run.

Worth noting the pre-existing code only looked bounded: it read the whole file and then used line 0. So this is not a regression #846 introduces, it is an existing cost neither the old code nor #846 avoids.

One semantic change here: readFirstEntry now returns only session entries rather than whatever line 0 parsed to. The caller's first.type !== 'session' guard becomes redundant but is left in place rather than widening the diff.

Acceptance criteria

• OMP title-first JSONL transcripts are discovered
• Blank lines before the session record do not prevent discovery
• Message-only files with no session record remain excluded
• Pi and OMP retain identical shared discovery behaviour

Testing

Fixture tests in tests/providers/pi.test.ts reproducing a title-slot-first OMP transcript, confirmed failing before the change. Also covered: blank-line tolerance, malformed leading JSON handled without throwing, message-only exclusion, a session record placed beyond the scan bound (which must not be found, proving the scan is bounded rather than full-file), and a large message-only file completing well inside a generous time ceiling.

Full suite 2471 passed, tsc --noEmit clean.

Credit

@jbspeakr reported #845 and traced the format to its upstream origin (can1357/oh-my-pi@0ce330ab, 2026-06-27), which is what made this straightforward to scope.

[iamtoruk]

Landing this for the one difference you flagged, now that #846 has merged the correctness fix. Rebased onto current main and reduced to exactly the perf delta: discovery streams through readSessionLines and stops after MAX_HEADER_LINES_SCANNED (20) leading lines, instead of #846's readSessionFile which reads each transcript in full (up to the 128MB cap) before looking at line 0. On the discovery hot path (walked for every file on every run), that turns a message-only file with no session record from a full read into a 20-line scan.

One reconciliation on top of your branch: the auto-merge reverted #846's readFirstEntry → readSessionEntry rename, so I re-applied it (and dropped the now-redundant first.type !== 'session' caller guard, since readSessionEntry only ever returns a session entry or null). Net effect on main is purely "bounded streamed scan," with #846's naming preserved.

Verified on the merged tree: tsc clean, omp + pi provider suites 37/37 (both #846's title-first omp case and your bounded-scan cases, including the "session record beyond the 20-line bound is not found" test that proves the scan is bounded rather than full-file), full suite green apart from the two pre-existing parser.test.ts failures. The 20-line bound is safe for the real OMP/Pi format (title slot + header at the top). Thanks for the analysis on the read-cost difference — it was the right call to keep it. Merging.