Record edits owed at the next wrap-up: friction annotations, a twice-wrong paragraph, and two sprint-plan corrections

[topij]

Bookkeeping, deliberately deferred. Severity L.

Three entries in docs/kit-friction-log.md were routed to the tracker on 2026-09-09 after the commit that would normally have annotated them had already merged (e698ec4). Their text still reads as parked, so triage-friction-log would re-file work that is already filed — the exact hazard the routing-annotation convention exists to prevent.

What to annotate

Entry (dated heading)	Opening words	Where it went
2026-09-09	"A lens imposed its own timeout on make test and killed its run…"	occurrence on #578
2026-09-07	"Lenses end their turn on a progress update while a background make test runs."	same occurrence on #578
2026-09-09	"Desktop hook testing looped on app mode while the task used another config."	#721

The first two share one comment because they were parked on the same open question — whether the lens had a foreground alternative — and one occurrence answered it for both. Their park conditions are discharged; say so rather than leaving them reading as still accumulating.

Locate them by their opening words rather than by line number; the file moves.

Why this was deferred rather than done

Annotating three entries is a repository change, so it needs a branch, a pull request and a panel. On the same day, the wrap-up record that carries these entries took a full panel, two delta passes and a second full panel, with each of the two delta passes finding a defect introduced by the previous round's fix. Spending that cycle on three annotation lines was judged out of proportion, and the operator agreed. #209 and #666 are where that cost belongs as a problem.

The trade is explicit: the annotation debt is cheap to carry for one session and is recorded here so the next session sees it in the backlog, rather than being carried in prose nothing re-reads.

Done when

The three entries name where each went, in the file's existing routing-annotation form, in an ordinary wrap-up commit. No separate pull request is wanted for this.

https://claude.ai/code/session_01936YzSmyhWE9xmeHWYGsn1

[topij]

Scope widened 2026-09-09, on the operator's decisions later the same day. Retitled to match. All of this is one ordinary wrap-up commit — still no separate pull request.

Everything below is a repository edit that was deliberately not made when the decision was taken, because each would otherwise buy its own branch, pull request and panel. Batching them costs one cycle instead of four.

4. Delete the /tmp expiry mechanism paragraph from the handoff — do not correct it again

docs/kit-handoff.md's continuity paragraph explains why the temporary trees expire, citing com.apple.tmp_cleaner, daily_clean_tmps_days="3" and -atime +3 -mtime +3 -ctime +3.

Verified 2026-09-09: both trees are gone.

MISSING /private/tmp/adk-adopt-field-20260905-5mfj1st8/fixture
MISSING /private/tmp/adk-adopt-continuation-20260906-AFeElK/kit-source

The documented rule accounts for kit-source (last stamped 2026-09-06) and does not account for the fixture, stamped 2026-09-07 and therefore two days old, inside the three-day window. So the mechanism as written is not what removed both.

That paragraph has now been wrong three times, in three different ways: d72bac4 asserted a reboot ended the route with no command behind it; ff248b1 replaced that with a claim sourced from an ls on a path that does not exist, reading a failure as evidence of absence; and this. Per Keep the record small — a claim that has needed repair twice gets cut to what you can stand behind — delete the mechanism, do not write a fourth explanation.

Keep the instruction, which worked. "Check both trees first, and stop rather than rebuild if either is gone" is independent of any mechanism, and a session following it today would have stopped correctly. Keep the stat-it-yourself line too.

5. Two corrections to saved_plans/codex-parity-plan_2026-08-23.md

That file is the maintained sprint status and is now stale on two points:

• Phase 5, delivery item 5 describes continuation work in the two temporary trees above. Both are gone and the standing instruction is to stop rather than rebuild, so the item cannot proceed as written. Record that it is blocked pending an operator decision on whether to rebuild from the retained saved_plans/ baselines or change approach — and record that item 6, the cs-toolkit replay, does not depend on those trees and is the phase's exit test.
• #585 moves earlier, out of Phase 6. The decision and its evidence are on #585.

6. Point the handoff's ▶ Next: at the cs-toolkit replay

The current ▶ Next: says several threads are live with none obviously first. That was true when written and is not now: item 6 is unblocked, is the exit test, and needs its own exact operator decision before it starts.

Why the delay is safe rather than merely cheap

The handoff's existing guard does the load-bearing work in the meantime: a session that reads it will check both trees, find them missing, and stop rather than rebuild. The risk of deferring is therefore not that a session does the wrong thing — it is that it stops without knowing what to do instead, which this issue answers from the backlog /session-start already reads.