2026-07-14 (M1, itd-89/spc-4) — Transcript capture is wired to `SessionEnd`, NOT the

2026-07-14 (M1, itd-89/spc-4) — Transcript capture is wired to SessionEnd, NOT the Stop the plan specified. The plan's letter was wrong on a matter of harness fact: Stop fires once per assistant TURN, and Claude Code's transcript file grows through a session, so a Stop-wired capture stores a fresh, larger superset every turn — proven by live test (one session, 4 turns → 4 records; a 100-turn session → 100 records and O(N²) bytes). history.Capture's sha256 dedup only collapses byte-IDENTICAL re-captures, which never happens on a live transcript, so the plan's "re-capture is idempotent" acceptance is false under Stop. SessionEnd fires once at termination and by contract ignores exit code + stdout — a perfect fit for a fail-closed, non-blocking side-effect hook. Verified against the harness docs (code.claude.com/docs/en/hooks). Accepted cost, recorded not hidden: SessionEnd does not fire on a hard crash/SIGKILL, so an uncleanly-killed session is not captured; the Stop-with-session_id-dedup alternative that would recover that case needs a change to shipped core dedup semantics and is deferred. This is the M1 deviation the loop is required to surface.