feat: parse full response_item transcript coverage (#13)

[ayushsingh82]

Summary

Addresses part 1 of #13 (the transcript-coverage half; the per-project tag override half is a separate, larger change and not included here).

parseTranscript() in src/services/transcript.ts only recognized response_item entries with payload.role === "assistant", and only extracted output_text content blocks. Recent Codex JSONL transcripts also carry, per #13:

• user messages as response_item.payload.type === "message" with role: "user" and input_text content blocks — previously dropped entirely
• assistant messages using text blocks in addition to output_text — text blocks were previously dropped
• function_call entries (tool invocations) — previously dropped entirely
• function_call_output entries (tool results) — previously dropped entirely

On a transcript that's mostly response_item-shaped rather than legacy event_msg-shaped, automatic capture could miss user intent and tool context.

Changes

• parseTranscript() now extracts all of the above. Tool calls/results are recorded as bounded, truncated "tool"-role entries (capped at 500 chars with a truncated, N more chars marker) rather than raw unbounded output, per the issue's ask to filter/truncate tool context.
• Since a rollout file can log the same turn in both the legacy event_msg shape and the current response_item shape, entries are now deduped when an identical (role, content) pair recurs within a small line window — so the same turn isn't captured twice.
• signals.ts's groupEntriesIntoTurns() now carries "tool" entries into a turn's captured content, but doesn't scan them for signal keywords — bounded tool output is noisy and shouldn't itself trigger a capture.
• build.mjs now also emits transcript.js as a standalone dist/services/transcript.js bundle (same pattern already used for session.js/tags.js/resultMerge.js), so this could be tested against real JSONL fixtures instead of only through the hook binaries.

Test plan

• npm run typecheck — no new errors introduced by this change (4 pre-existing errors on main, unrelated to this file, are already being fixed in fix: don't lose capture delta on failed upload; restore clean typecheck #36)
• npm test — 74/74 pass, including 6 new fixture-based tests in test/unit.mjs under "Codex transcript parsing":
  • extracts response_item user messages from input_text blocks
  • extracts response_item assistant messages from output_text and text blocks
  • captures function_call/function_call_output as bounded tool entries, verifying truncation
  • does not double-capture a turn logged as both event_msg and response_item
  • legacy event_msg-only transcripts still parse unchanged (regression guard)
  • an identical message repeated far apart (outside the dedup window) is still kept as two entries

[ayushsingh82]

@Dhravya parses full response_item transcript coverage (user/tool messages that were previously dropped). Please take a look.