fix(llc): resolve the run transcript from the adapter state file (#13614)

[mrveiss]

Closes #13614.

Thinking Path

A claude_code run executed real work end to end — 5 turns, real tool calls,
is_error: false — and its 37 KB transcript sat on disk while
llc_run_replay_logs.output_text stayed empty and recorded_events NULL.

Two places derive the same path from different inputs:

adapter    _output_path(dir, agent, "0/<session>")   -> ..._0_<session>.jsonl
adapter    return f"{proc.pid}/{session_id}"         -> "<pid>/<session>"
recorder   _output_path(dir, agent, external_run_id) -> ..._<pid>_<session>.jsonl   <- never exists

The placeholder is not an oversight that can simply be corrected: the file is
the child's stdout, so it must be created before Popen, and the pid does
not exist until after. Any fix that keeps two derivations has to keep them in
step forever.

The adapter already records the truth. Its state file — written after spawn,
keyed by the id it returns — carries the real output_file. So the fix is to
read it and delete the second derivation entirely.

What Changed

heartbeat_scheduler.py gains _resolve_adapter_output_file, which reads the
adapter's state file for the returned run id and takes output_file from it.

It falls back to the computed path when the state file is missing or unreadable
— runs predating the state file still resolve as well as they used to, and the
caller already reports a missing transcript rather than silently recording an
empty one.

Verification

Reproduced the defect and the fix against a stand-in that mirrors the adapter
exactly — placeholder-named file, pid-bearing returned id, state file written:

transcript on disk : llc_agent_agent7_0_<session>.jsonl
OLD resolver       : llc_agent_agent7_4242_<session>.jsonl  exists=False   <- the bug
NEW resolver       : llc_agent_agent7_0_<session>.jsonl     exists=True
no state file      : falls back to the computed path

A regression test drives _record_run_for_replay itself, with the file
named by the placeholder and a state file keyed by the returned id. It asserts
up front that the computed path does not exist — otherwise the test would
pass through the fallback and prove nothing — then asserts the transcript is
attached and recorded_events is non-empty.

The existing H1 test still passes, through the fallback: it creates the file
at the computed path with no state file, which is now the fallback's exact
case. So both paths are covered, and the older behaviour is preserved rather
than replaced.

Not addressed

The 0/ placeholder remains in the adapter. Renaming the file post-spawn would
remove the oddity but adds a failure mode (a rename that half-completes) to a
path that currently cannot lose data. The state file is the contract; this makes
the recorder use it.

Model Used

Claude Opus 5

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!