chore: update handoff — the fifth sweep, and a claim wrong in both directions

[topij]

End-of-session wrap-up for 2026-07-29. Handoff + friction inbox only; no code changes.

Handoff

New ## Latest session block for the fifth inbox graduation (#151, merged 494b9eb). The session's durable result: an operator's narrow remark widened into a stronger claim and published as operator-confirmed, then used to file a structurally-never-reviews verdict onto #45 — the issue whose whole subject is that such a verdict cannot be made from outside. The fix over-corrected; a second panel round caught that; two independent re-derivations of the number disagreed, so no count was published.

Two panel rounds on this PR, and what they cost

Round 1 found nine defects in this commit. The correction of those nine introduced three more, which round 2 found. All twelve are now fixed. This section is longer than the change because that is the honest shape of it.

Round 1 — the worst of the nine. The wrap-up called #151 "the third sweep" in a heading, the Last updated line, the commit subject and the PR title — while the same block said "Fifth sweep overall" three lines below, and while #151's merged body records "Third sweep" → fifth as one of its own review corrections. Both lenses caught it. It is now a friction entry with a sharper remedy than #149's: the handoff was drafted from the session's memory rather than from the merged artifact, so the corrected value never entered the drafting at all.

The other eight: a stale inbox line count on two surfaces; a closing-keyword verification claiming "0 matches" while a later paragraph said two remained, over a character count that reconciled to nothing; a miscounted set of added inbox entries; the #73 bullet naming the previous session's instance instead of the new one this sweep created; the silent-PR run stated as twelve when #151 makes it thirteen; an unannotated superseded ▶ Next:; a blank line rendering the "Open" list loose; and "three unrecorded observations" that were two.

Round 2 — three defects in the fix itself.

1. The entry count, corrected into a different wrong value. The original said "four added" while naming the config/dev-model.yaml's tracker block is unstamped placeholder and names the wrong backend #121 entry as one of them. The fix said "three added" — also wrong, and published as a correction. The truth is 4 added by this commit, 5 total under the heading, because the fix round itself added a fifth entry and never re-derived. From git: 43 insertions, 0 deletions on the friction log; 1 entry under that heading on main, 5 at HEAD.
2. "No cross-reference lost" was false, and this same document disproved it three paragraphs later. The archive move is verbatim — a lens re-ran the script and confirmed byte-identity — but it created a new archive_plan_sessions must warn on relative cross-references — both inside moved blocks and into them #73 instance: a relative "the 287-line figure above" now dangles in kit-handoff-history.md while its referent stayed in the live handoff.
3. The handoff Theme's tally went stale inside the commit documenting the extra rounds — it said "two panel rounds, four isolated lenses" while this PR was adding a third and fourth. Now four rounds, seven lenses.

Archive sweep

archive_plan_sessions.py with the default --keep 6 was a no-op — six blocks already exceeded the 400-line budget. Re-run with --keep 4: two blocks moved, 465 → 322 plan lines (325 after corrections). Verbatim, independently re-run by a lens — but not cross-reference-clean, per round 2 item 2 above.

Friction inbox

Five entries under ## 2026-07-29 (post-sweep); four added by this commit, the fifth (#121) landed by #151.

• H — an operator's remark widened into a stronger claim and published as operator-confirmed on five surfaces. Distinct from When a claim is corrected, enumerate every surface it was published to at that moment #149: that asks a correction to reach every surface; this asks the claim not to exceed its source.
• M — correcting a wrong number with a precise one failed twice because the number is not recoverable. Occurrence data for Nothing distinguishes a bot that structurally never reviews from one that is merely pending #45.
• M — a check that errored reported a pass, inside the step guarding Build the closing-keyword guard: flag every match on every surface, including the squash message #71; a later run scanned a zero-byte surface and also read clean; and the rewritten scan's own pattern dropped the code-span case the rule names explicitly.
• M — the wrap-up reinstating a merged correction.

The inbox is 179/150 — over budget the same day it was swept (203 → 136 by the sweep, then back up by this wrap-up's own entries). Not swept inline: the wrap-up contract reserves graduation for triage-friction-log, which needs tracker writes and operator approval. Noted in the handoff as an argument that #6 (vendor the engine) is closer to urgent than merely open — a sweep now buys about one session of headroom.

Closing-keyword verification

Pattern admits backticks, quotes, emphasis marks and brackets between keyword and reference, because the repo's rule says code spans do not protect.

surface	size	matches	triggers closing?
commit message	2,549 chars	0	yes
PR title	—	0	no
PR body	this text	0	yes
added diff lines	267 lines / 19,705 chars	2	no

Both diff matches are in kit-handoff-history.md, moved verbatim by the sweep script, and name issues 92 and 100 — both already CLOSED. File content does not trigger GitHub's closing behaviour; only commit messages and PR bodies do. A lens additionally confirmed the gerund form is not one of GitHub's accepted keywords, making them doubly inert. Left as the historical record.

A third match existed in an earlier revision, in newly-authored friction-log prose that quoted the construction while describing it. It is now written as prose — quoting it would plant a live pattern in an inbox entry a future sweep pastes into an issue body.

Where this stops

Stopping criterion: second class (a record that is read, not a gate, send path, or destructive operation) — act on every HIGH and on any regression at any severity. Done for all twelve.

This is not a "the last round found nothing" stop and is not claimed as one. Every round so far has found something, and the panel doctrine records that the termination condition may never arrive. What has changed is the target: the shipped artifacts — the handoff and the friction log — were independently re-derived clean by a lens (every count, the archive move, every issue state). The surviving defects have all been in the prose describing the change, which is the same finding this session filed twice.

Verification

make test — 503 passed. check_doc_budget — handoff 325/400 ✓, friction log 179/150 ⚠ (deliberate, above). kit_doctor — 26 unchanged, 0 differ. Both toolkit CI checks green.

Issues referenced above are context, not targets — none of them is being retired by this PR.