[HARDENING] The mumble discriminator fails its own must-detect gate; repairs stopped at the gate (PENDING-179)

The 2026-09-03 repair plan rested on one claim: that human_turns() is an
existing, controlled discriminator, and wiring it into four transcript
consumer sites was plumbing rather than classifier-building.

Tested against ground truth taken from the fool's own prompt text — not from
the function under test — the claim is false. 24 known mumbles, 41 known
non-mumbles, out of 65 transcripts:

  must-detect    22/24  two mumbles read as human-attended
  must-not-flag  35/41  and one "failure" is correct — b7e7eb39 is the
                        unattended session of 2026-08-31, which genuinely
                        has no human turn

human_turns() == 0 never meant "mumble". It means "nobody spoke", which is
equally true of an unattended real session. A mumble embeds the previous
session's text and so inherits its slash-command markers; it is excluded
only when the session it quoted happened to contain one. The two leaks are
exactly the two marker-free mumbles. Coincidence, not design.

No repair was made at any site, and no replacement discriminator was built.
Three controls passed and a fourth broke: all three test the question the
function was built for, none could see the question it was being reused for.
A successor written now inherits whatever made the first set look sufficient.

Also here, per the 2026-09-03 handover:
  - step 0 preservation ran: 76 transcripts, read-back PASS, 11 of them
    already pruned at source and surviving only in the archive
  - gate 2b: the composition claim in PENDING-178 stands (11 mumbles,
    ~a fifth, on 08-31). Two terms do not close and are reported as an open
    disagreement, not a correction — the reconstruction is a demonstrated
    lower bound and -178 enumerated live
  - gate 2c: five sites in four files; -178's [HARDENING] scope holds
  - MEMORY.md record-only arithmetic correction: N-now 44 -> 65 measured,
    composition 41 real + 24 mumble added, direction reversed (it is rising,
    not shedding), stale :331 pointer corrected to :513

Filed as an item rather than a PENDING-178 addendum on PENDING-145's
mechanism: a ruling claims a number, so an addendum would be suppressed the
moment -178 is ruled. The undecided filing moratorium is disclosed inside
the item.

governance-drift-check.py:513 excluded on receipt and not examined.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X3L79vgAnt1x2kxvf23Qt7

claude/memory/session-ledger-2026-09-03.md

@@ -0,0 +1,128 @@
+---
+name: Session Ledger 2026-09-03
+description: Practice-of-return ledger maintained by /symmetria — returns, open horizons, recalibrations, authorization moves, sub-agent dialogues, bypasses.
+type: feedback
+---
+
+# Session Ledger — 2026-09-03
+
+> ⚠ Note on the file itself: this ledger was ABSENT when the second session of 2026-09-03 woke,
+> though a full session had already run and wrapped 2.2 h earlier. Symmetria was not init'd that
+> session (or was and did not write). The wrap record used ledger-shaped language ("Instruments: 2
+> run · K = 0") with no ledger behind it. **A ledger that only exists when someone remembers to
+> invoke it is not a record of returns; it is a record of invocations.** Kin to PENDING-168.
+
+## Returns
+
+- **2026-09-03T~14:0x — INHERITED CLAIM FALSIFIED AT THE WAKE, BY RUNNING IT RATHER THAN RELAYING IT.**
+  The prior wrap recorded the `wake-digest.py` selftest as failing and predicted it would *"fail on
+  every run from now on"* (0 of 13 sample transcripts wrapped; all 13 Tarbuckle mumbles). Re-run at
+  this wake: **5 of 13, SELFTEST PASS.** The prediction was falsified in 2.2 hours. Corrected
+  statement of the defect: the control's verdict is a **function of the mumble rate in a rolling
+  13-transcript window**, so it is *intermittent*, not permanent — a harder failure to notice than
+  the one recorded, because it will read clean on some runs and dirty on others with no change in
+  what it is measuring. Return caught by the standing discipline *prove the instrument before
+  trusting a clean line* — applied here to a DIRTY inherited line, which is the same rule run
+  backwards and was not obvious.
+
+- **2026-09-03T~14:0x — measured N-now rather than relaying it.** MEMORY.md carries 44 (2026-08-31);
+  the prior wrap's addendum carries 52. Actual, by the trial's own method: **65**. Then went one step
+  past the number the freeze obliges me to report, and censused its composition with the controlled
+  discriminator (`human_turns`): **31 of 65 are mumbles (48%), 34 real, 0 unreadable.** The bare
+  count would have been true and useless; the composition is the finding.
+
+
+- **2026-09-03T~15:0x — THE QUERY THAT FOUND NOTHING, AND THE VACUOUS PASS IT PRODUCED.**
+  Building 2a's ground truth I searched for the Tarbuckle prompt signature `"You are Tarbuckle.
+  Your character..."` — taken from `tarbuckle-invoke.py`, the file I had just read. It matched
+  **0 of 65** transcripts, and the gate duly reported **`PASS (0/0)`**. A green light with an
+  empty denominator. Caught by the standing rule *a null search is evidence about the QUERY*;
+  opening a transcript showed the real signature is `"You are writing ONE line as Tarbuckle"` — a
+  DIFFERENT fool surface (`tarbuckle-mumble.py`/`-wrap`/`-seam`, three scripts the inherited site
+  census never named). ⚠ **The flag that fired is §3's *a search query shaped by what the session
+  wants to find*.** Had I accepted the pass, I would have reported the discriminator sound and
+  wired a broken predicate into three more sites.
+
+- **2026-09-03T~15:1x — FRAME-INHERITANCE, and it is the session's whole finding.**
+  `human_turns()` carries three passing controls. All three test the question it was built for
+  ("did an executor run unattended?"). The plan reused it for a different question ("is this a
+  mumble?") and inherited the controls' authority across that gap. Re-run at the scope of the
+  extension: **22/24 must-detect, 35/41 must-not-flag — FAIL both ways.** The exclusion works only
+  when a mumble happens to quote a slash command; the 2 leaks are exactly the 2 marker-free
+  mumbles. ⚠ **`0` never meant "mumble" — it means "nobody spoke", which is also true of a
+  genuinely unattended session (`b7e7eb39`, PENDING-172).**
+
+## What held  [appended]
+
+- **Stopped at the gate instead of building the replacement.** The handover's instruction and the
+  reasoning behind it were both honoured: three controls passed and a fourth broke, so the finding
+  is about the control SET. Writing a successor discriminator now would inherit whatever made the
+  first set look sufficient. Sites 2-4 stay shut.
+- **Ran 2b and 2c anyway**, because 2a's failure does not gate them and they answer different
+  questions. 2b's composition term closed; 2c's scope held.
+- **Did not assert 2b's gap against -178.** My reconstruction disagrees by 4 files and is a
+  demonstrated lower bound (preservation ran twice only). Reported as an open disagreement with
+  -178's method named as the better-positioned one, rather than as a correction.
+- **Narrowed the census three times and said so.** Repeated narrowing can end by confirming the
+  five sites it was built around; the negative control and the explicit blind-spot statement are
+  what keep that honest.
+
+## Authorization moves  [appended]
+
+- **PENDING-179 FILED** — as an item, not a -178 addendum, on PENDING-145's demonstrated mechanism
+  (a ruling claims a NUMBER; an addendum under a to-be-ruled number is suppressed on arrival).
+  ⚠ The undecided moratorium is disclosed inside the item, with yesterday's contrary reasoning
+  preserved rather than overridden.
+- **MEMORY.md record-only arithmetic correction applied** (authorized by the handover, 2b): N-now
+  44 -> **65 measured**, composition **41 real + 24 mumble** added, the "shedding faster than it
+  gains" direction reversed, and the stale `governance-drift-check.py:331` pointer corrected to
+  `:513` (`:331` is a register-check control, verified by reading both lines).
+- **NOT done, deliberately:** no repair at any site; `governance-drift-check.py:513` not examined,
+  excluded on receipt; no replacement discriminator; nothing committed (the wrap owns that).
+
+
+- Did not answer the inherited literal question at the wake despite it being cheap to answer
+  (`git log -p` on MEMORY.md). The wake's rule is to hold it open. Held.
+- Reported the thread-query null as a null, in one clause, rather than dressing six generic
+  term-matches as findings.
+
+## Open horizons
+
+- **The steward-directed first work:** wire `human_turns()` into the transcript-as-session consumer
+  sites. Positive control FIRST (yesterday's cost of skipping one: a false loss report within the
+  hour). `tarbuckle-invoke.py:40` is SUSPECTED, unverified — verify before touching.
+- ⚠ **The 84 trigger is ~19 transcripts away and roughly half the counter is chatter.** The ladder
+  freeze (REVIEWED-123) is measuring against a counter PENDING-178 says is mis-united. Time pressure
+  is now real and was not when -178 was filed.
+- Steward owes: `~/.claude/agents` under version control (before anything else edits it) · the
+  moratorium decision · `RE_ID` `[FIX]` · PENDING-171/-177/-178/-160/-168 · the §5 regrade.
+- Parked worker `acaabadf` — stopped, still carrying `--reply-on-resume`; respawn behaviour not
+  established. Untouched.
+
+## Confidence to recalibrate
+
+- **Verified this session (ran it):** selftest 86→87 passing, 5 of 13 (`--selftest`) · N-now = 65
+  (glob, the trial's own method) · mumble census 31/34/0 (`human_turns` over all 65) · dotfiles
+  local + gitea + github all at `e5db321`, only dirt is this wake's own `thread-query-log.jsonl`
+  append · governance drift 0 · pointers 407/0/0 · today's ledger absent before now.
+- **Inherited, NOT re-verified:** the four-site census in ADDENDUM 1 (grep-derived, read not re-run)
+  · that `human_turns` is the right discriminator at every site (proven at one) · that
+  `tarbuckle-invoke.py:40` is affected at all.
+- **Post-compression note:** context was cleared deliberately between the two sessions of 2026-09-03.
+  Everything above marked "inherited" rests on the written record, not on continuity of working
+  memory.
+
+## Authorization moves
+
+- Reaffirmed the prior session's own correction: the fix is **`[FIX]`, not `[PROPOSAL]`** — the
+  control's label already states its predicate ("a real session reads as WRAPPED end-to-end"), so
+  restoring that population repairs it against its existing spec. ⚠ Does **not** extend to changing
+  what the `>=84` trigger *means* — that stays PENDING-178 / REVIEWED-123, steward's.
+
+## Sub-agent dialogues
+
+*(none)*
+
+## Bypasses
+
+*(none)*