skill-harvest: two proposals earned by failures inside this wrap

Both patch /wrap-up steps, and both are filed as PROPOSAL rather than taken
through the FIX lane for the same reason: each edits a step while that step is
being executed by the session proposing it, which is the configuration where the
classification test is least trustworthy. When in doubt, propose.

  - §3's MEMORY.md rotation is a two-file write with no atomicity, and it
    half-applied today, leaving the prior Active Session in both files.
  - §7.5's required daily-note shape is asserted in two places and checked in
    none. Three required sections were missing today, including ## Corrections
    — the section REVIEWED-126 added because a format with a slot for insights
    and none for errors under-records errors.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RmFYCUeAaPqbpJMj6uGokk

claude/memory/skill-harvest-register.md

@@ -575,3 +575,14 @@ it, which is the argument for the step.
 ⚠ **Why this is a `[PROPOSAL]` and not the FIX lane.** It adds a *required check* to a governed instrument rather than changing what that instrument records, and the first clause of the classification test — "what the executor may do without asking" — is arguable either way. The lane's own instruction is *when in doubt, propose*. It also touches `governance-drift-check.py`, which REVIEWED-95's falsifier leans on; the ladder trial is frozen under REVIEWED-123 and **adding a reporting section is not obviously outside that freeze's spirit even though it is outside its letter.** That doubt alone routes it here.
 
 ⚠ **The self-referential note, since it is the point.** This proposal exists because the executor built (d) as a recall-dependent CLI **on the same day, in the same item, as the register line recording that recall-dependent capabilities fire at ~0%.** The measurement was in front of me and the build ignored it. That is not a discipline failure to be fixed with more care — it is the routing table's own thesis, demonstrated at its own expense.
+
+## Proposed 2026-08-27 wrap — awaiting steward
+
+| # | kind | target | proposal | firing moment | routing |
+|---|---|---|---|---|---|
+| — | patch | `/wrap-up` §3 | **The MEMORY.md rotation is a two-file write with no atomicity, and it half-applied today.** §3 demotes the prior Active Session into `MEMORY-reference.md` *and* rewrites `MEMORY.md`. This wrap's script wrote the reference file, then raised on the next statement, so **the old Active Session existed in BOTH files** — the exact "never leave two Active Session entries" hazard the step exists to prevent, arrived at from the other side. Caught by reading back, not by the step. Proposed: §3 states the order (**write MEMORY.md first, demote second**, so a mid-failure leaves one copy rather than two) and ends with a three-line assertion — exactly one `## Active Session`, zero occurrences of the prior session's pointer in `MEMORY.md`, exactly one demotion marker in the reference file. ⚠ **Honest limit:** this makes a partial write *detectable*, not impossible; two files cannot be written atomically here. | **at every `/wrap-up` §3** — already a ritual step | named step in an existing ritual (strong tier) |
+| — | patch | `/wrap-up` §7.5 + `daybook-ensure.py` | **The daily note's required shape is asserted in two places and checked in none.** §7.5 lists six required `##` sections and warns *in its own text* that the skeleton is a second copy which "will drift". Today it had drifted in effect: the note carried `The day in short` and `Open / next` but **not** `Decisions taken`, `Insights and exchanges worth keeping`, or `Corrections` — and I only found out by grepping the headings against the list by hand. ⚠ **`## Corrections` is the one that matters**: REVIEWED-126 added it precisely because a format with a slot for insights and none for errors under-records errors, and it was the missing section on a day with roughly a dozen corrections. Proposed: a check that reads the day's note and reports missing required sections — mechanically, from the skeleton's own `SKELETON` constant so the two copies cannot disagree. | **at every `/wrap-up` §7.5**, or better, at the `Stop` hook that already fires the daybook cue | mechanical detector (strongest tier) — the cue hook already runs on a timer and already reads this file |
+
+⚠ **Both filed as `[PROPOSAL]`, not taken through the FIX lane, and the reason is the same for both.** Each edits a `/wrap-up` step **while that step is being executed by the session proposing it**, which is the one configuration where the classification test is least trustworthy — the executor is both author and beneficiary of the judgement that it needs no authorization. The lane's own instruction is *when in doubt, propose*, and self-modification during execution is the definition of doubt.
+
+⚠ **Both were earned by failures in THIS wrap**, not observed in the abstract: the half-applied rotation actually happened, and the missing sections were actually missing. Neither is a speculative improvement.