Make the posted thinking trace opt-in and bounded

[asavs]

The model's thinking trace was prefixed onto every posted review body unconditionally.

Measured across every review this bot has posted — 171 reviews spanning 2026-07-13 to 2026-07-28, across asavs/mog-template and Goonistotle/third-person-multiplayer:

reviews carrying a trace	169 of 171
aggregate trace share	74% (833,421 of 1,116,274 bytes)
median per-review share	69% (p10 34%, p90 91%)
median trace	3,622 bytes
largest trace	29,809 bytes, in a 31,477-byte review (94.7%)

So this is not a fix for one pathological review: roughly three quarters of every byte this bot has ever posted is trace rather than review. Those byte figures are exact.

The trace is not junk, and this PR is not a judgement on its quality. Bucketing the 1,569 trace section headers across those 169 traces by keyword: ~60% analysis or verification of the code, ~18% crafting and finalizing the review, ~3% tool or permission friction. The model is largely reasoning about the code, not padding or flailing. That split is a proxy and weaker evidence than the byte counts above — it classifies the model's own section titles, not the paragraphs beneath them, so a header reading "Verifying Error Handling" is not proof the body did so.

The case for taking it out of the posted body is about the channel, not the content:

• it is already captured in three places that cost no reader anything — the runtime sidecar, research artifacts under consent, and agy's own transcripts;
• <details> collapses it only in a browser. The GitHub API returns raw markdown, so every other reader ingests all of it — including this daemon when it reads back its own prior review;
• it outweighs the review it is attached to by roughly 3:1.

The change

• REVIEWER_POST_REVIEW_TRACE (default 0) governs whether the sidecar trace is prefixed onto the posted body.
• REVIEWER_TRACE_MAX_BYTES (default 20000) bounds it when enabled, cutting on a line boundary and marking what it dropped. GitHub rejects review bodies over 65,536 characters outright, so an uncapped trace did not degrade a review — it failed to post it. At 20,000 the cap is a safety valve rather than routine truncation: only 2 of the 169 observed traces would have been cut.

Capture is unchanged. The trace still reaches the runtime sidecar, research artifacts under consent, and agy's own transcripts. Only the posted body changes.

Also fixes review_subject_from_body, which now skips a leading <details> block. The trace writes bold paragraphs rather than headings, so the first heading it found was the one after the trace — in production the prior-review summary the daemon fed itself was resolving to Reference Links, the review's closing section.

Not gated by the flag: the branch that collapses trace-like prose the model wrote into its own body. That text is the model's output rather than something the daemon appended, so only its formatting changes.

Verification

650 fixture assertions, green. The end-to-end test asserts on the JSON payload the stubbed curl actually receives (--data → jq -r '.body'), running the loop under both flag settings.

Proven non-vacuous: flipping the production default from 0 to 1 turns the flag-off assertion red —

POST_REVIEW_TRACE="${REVIEWER_POST_REVIEW_TRACE:-1}"
not ok - posted body omits the sidecar trace when the flag is unset

— and restoring it returns the suite to 650 green.

One note for reviewers: a dry run cannot verify this. write_dry_run_artifact is passed the raw model response, while the trace is prefixed later into the posted body, so a dry-run artifact shows no trace regardless of the flag. I initially "verified" it that way and it proved nothing. The fixture plants its sidecar the way a real run produces one (a stubbed brain transcript), because run_agy_review deletes and regenerates a directly-planted thinking.trace.

Note: conflicts with #177 on the pinned EXPECTED_ASSERTIONS line — both branches bump it from 633. Trivial to resolve, but guaranteed for whichever merges second.