Close the changelog gap and retire the swe-bench evidence

[shyamsridhar123]

Two documentation defects, both of the "advertised in one place, true nowhere" class this repo keeps finding.

1. Four merges shipped undocumented

The last merge to touch CHANGELOG.md was #198, and the highest issue referenced in [Unreleased] was #195. Everything since shipped with no entry:

Merge	Entry?
#209 — hybrid mission across two backends (#86)	missing
#208 — eval-gate instrument validity	missing
#207 — contract restores every file it names	missing
#205 — timeout_secs enforced, truncate panic	missing

scripts/release_drift.py did not catch this and structurally cannot: it verifies entries exist, not that recent merges are among them. It reported OK -- 102 commit(s) since 0.5.0, 28 [Unreleased] entr(ies) while the four most recent merges were all absent. Worth noting as a gap in the drift check itself; not fixed here.

Entries added for all four. [Unreleased] now reports 38.

2. The README presented a retired benchmark as live evidence

The swe-bench-style result sat as a row in the results table, and a limitations bullet described the gate in the present tense: "has no headroom left" — language about a gate still in use. The benchmark is out of scope for this project and the measurement has not been refreshed since 2026-07-03.

Marked retired in the results table, the hypothesis table in docs/intent-to-outcome.md, and the limitations bullet, which now says the gate is withdrawn and points at the instrument-validity rule from #208 that governs Tier 3 instead.

The numbers stay. The eval ran and the result is real. Deleting it would be the dishonest option; the fix is to stop presenting it as current.

What was deliberately left alone

Methodology citations are not touched. docs/autonomous-workflows.md cites the swe-bench discipline — confirm the check fails, fix, confirm it passes — which is the origin of Phoenix's failure-first gate, and src/accept.rs says the same in a comment. That lineage is not a stale claim about our results, and erasing it would erase where the idea came from.

BUILDLOG.md, the CHANGELOG's own history, and research/autonomous-workflows-research.md are historical records, left as written. handoff.md's mention is a live operational note about pytest collection, still accurate.

No eval code or eval results were modified.

The guard

Retiring a claim in prose is not durable on its own: the next person editing the results table has nothing telling them the row is historical, and a revert restores a current-tense claim with nothing to catch it.

tests/test_swebench_evidence_is_retired.py pins it, the same way test_okf_claim_discloses_its_limit.py pins the OKF cost/quality disclosure. It asserts only that the result is never presented as current evidence — and it also fails if the row disappears entirely, because retirement keeps the record rather than deleting it. Table-row scoped so methodology prose cannot trip it. The detector carries its own unit tests, without which the file would pass trivially if the mentions were simply deleted.

Verified RED against the pre-retirement docs (3 failed) and GREEN with them (6 passed), which is what the acceptance contract now points at.

Verification

49 tests across the doc-pinning suites — README accept claim, release drift, version consistency, OKF claim disclosure, scoreboard, scoreboard validity, auto-merge gate, ralph example gate — green.

One thing worth flagging in review: an earlier commit on this branch staged the two doc files at their pre-retirement revisions, because verifying the guard goes RED used git checkout origin/main -- <paths>, which writes the index as well as the working tree. 9b26553 restores them. The final diff against main is what matters and is correct; the intermediate commit is not.