docs(receipt): Stage 2 council dispositions, full-population census transitions

The single seat returned REJECT with four MAJORs; every finding was checked
at source. Two were real planner defects (fixed in c8c105c); the other
four were evidence-class findings, answered here with artefacts rather
than prose: the golden gate is marked NOT MET AS WRITTEN and amended to
projected equality checked between named commits; the census attribution
now rests on all 5,434 questions through both arms (identical 1,844 /
crash 2,349 / whole-question-phrase 1,227 / different 14; zero
regressions), with the mechanism read from the frozen escape function
itself (1,217 of 1,227) and the instrument difference from run_census
stated; the gold parity and rescued-crasher claims are counted per item;
the studyloop suite attribution rests on a matched control at 80ee57b in
the same sandbox (299 of 305 failing ids in common, identical exception
profiles). The four census texts affected by the F2 fix keep their
recorded outcomes, so the Stage 2 receipts stand exactly.

Author: NetDevAutomate

.secrets.baseline

@@ -24,47 +24,63 @@ An empty result is never silent: `plan="none"` plus a note says nothing was sear
 |---|---|---|
 | crashes through the real MCP tool | 0/91 | **0/91** (frozen control still 42/91) |
 | crashes through the CLI | 0/91 | **0/91** (was 42/91 + zero rows on the other 49) |
-| CLI ≡ MCP, ordered session lists | 91/91 | **91/91**, paired delta exactly 0 |
+| CLI ≡ MCP, ordered session lists | 91/91 | **91/91** — counted directly over the per-item `ranked` arrays (`stage2-census-transitions.json` → `gold_counters`), not inferred from equal aggregates |
 | K-stratum paired delta vs frozen | not negative | K **0.182 → 0.303**; P 0.034 → 0.034; R 0.103 → 0.138 |
-| planner golden | unchanged | **cases byte-identical**; row shape extended additively — see deviation |
+| planner golden | unchanged | **NOT MET AS WRITTEN — gate amended** (see below): equal after projecting `message_id` away, checked between `80ee57bb` and `d060d3f2` by `scripts/eval/stage2_council_evidence.py` |
 | macro recall@5 (reported, not gated) | — | **0.1066 → 0.1585**, paired +0.052, CI95 [+0.011, +0.100] (57 clusters) |
 | latency | — | mcp p50 20.8 ms / p95 92 ms (was 18.5 / 94.5) |
 
-Honest detail: of the 42 former crashers only **5** become hits at K=5. The crash class is dead;
+Honest detail: of the 42 former crashers only **5** become hits at K=5 (`frozen_crash_and_live_hit_at_5: 5`, counted per item; `moved_outside_crash_class: []`). The crash class is dead;
 those questions now have the same ranking problem as everything else. That is Stage 4's problem,
 and the K/P/R shape says so: K and R moved, P did not.
 
-## Deviation from the freeze wording — recorded, not hidden
+## Gate amendment — the golden gate was changed, not met
 
-The freeze said "golden file unchanged". The golden's `row_keys` gained `"message_id"` and each
-case result gained its id; every case's query, ordering and preview is byte-identical (verified
-by diff: five added lines, nothing removed). The gate's purpose — the explicit-syntax and simple
-cases must not move while the planner changes — held. Real message ids were a stated Stage 2
-requirement (the agent's interface had no citation handle and the census could not gate the tool
-without them), so the additive change is the requirement, not drift.
+The freeze said "golden file unchanged". That literal gate is **not met**: `row_keys` gained
+`"message_id"` and each case result gained its id (15 changed lines in the diff stat: the
+trailing-comma lines count as changed). The seat was right that labelling this "unchanged" was
+wrong. The gate is hereby **amended** to: *the golden is equal to its Stage 1 form after
+projecting `message_id` away* — checked between named commits `80ee57bb` and `d060d3f2` by
+`scripts/eval/stage2_council_evidence.py` (`golden_projection.equal_after_projecting_message_id_away:
+true`, `added_row_key: ["message_id"]`). Why the amendment rather than a waiver: real message ids
+were a Stage 2 requirement (the agent's interface had no citation handle; the census could not
+gate the tool without them), and the invariant the gate protected — explicit-syntax and simple
+cases do not move while the planner changes — is exactly what the projected equality checks.
 
 ## Census through the real tool — first time possible
 
 `stage2-census-mcp.json`: 5,434 eligible · hits **3,345** · hit@5 **0.6156** (Stage 1 frozen
 0.1305) · crash **0** (was 2,349) · vocabulary 257 · ranking 1,832 (untied 1,546) · tied hits 0.
 
-A jump that size was not accepted on sight. A 400-question paired probe (seed 11, both arms,
-same self-exclusion, session lists compared) attributes all of it to two shipped defects and
-finds **zero regressions**:
+A jump that size was not accepted on sight. The **full population** was re-run through both arms
+with identical self-exclusion and the ranked *session lists* compared per question
+(`stage2-census-transitions.json`, `per_question` carries every row; 1,834 s):
 
-| class | n | frozen own-session hits | live own-session hits |
+| class (frozen → live) | n | frozen own hit | live own hit |
 |---|---|---|---|
-| identical ranked lists (planner ran in both) | 140 | 54 | 54 |
-| frozen **crash** (the and/or/not class) | 189 | 0 | 129 |
-| frozen **empty**, live rows — **70 of 71 texts contain a `"`** | 71 | 0 | 57 |
-
-The second defect the gold set could not see: the shipped code treated any question containing
-a double-quote character as "explicitly quoted" and searched its *entire text* as one exact
-phrase — which matches nothing but the question itself. Learner turns quote error messages,
-file names and code constantly (71 of 400 = 18%). The planner now lifts quoted spans out as
-phrases and plans the rest.
-
-Sample hit@5 frozen 0.135 / live 0.600 — both within 0.02 of their full-census receipts.
+| identical ranked lists (planner ran in both) | 1,844 | 845 | 845 |
+| frozen **crash** (the and/or/not class) | 2,349 | 0 | 1,577 |
+| frozen **empty**, live rows | 1,227 | 0 | 1,016 |
+| different lists | 14 | 4 | 4 |
+| **regressions** (frozen hit, live miss) | **0** of 5,434 | | |
+
+Mechanism, read from the shipped code rather than inferred: for each frozen-empty question the
+Stage 1 replica's own `_frozen_escape_fts_query` was applied to the text — **1,217 of 1,227**
+produce a single double-quoted phrase spanning the whole question (`whole_question_as_one_phrase`;
+10 `other`). The shipped code treated any text containing a `"` as "explicitly quoted" and
+searched the *entire question* as one exact phrase, which matches nothing but the question
+itself. Learner turns quote error messages, file names and code constantly (22.6% of the
+population). The planner now lifts quoted spans out as phrases and plans the rest.
+
+Bound on the attribution: the transition table is a paired classification, not an ablation.
+The crash class and the whole-phrase class are disjoint by construction (a question is
+classified by what the frozen code did to it), so the counts above are exact for this corpus;
+they are not a claim about other corpora. One instrument difference, stated so the numbers are
+not mistaken for each other: the transition script excludes the question's **single** message
+id, while `run_census` also excludes its byte-identical re-asks within the same session; so the
+script's absolute rates (frozen 0.1562 / live 0.6334) run slightly above the receipts (0.1305 /
+0.6156). Both arms receive identical exclusions inside each instrument, so the classes, the
+mechanism and the zero-regression count stand; the receipts remain the official numbers.
 
 Per source (Stage 1 → Stage 2): claude_code 0.016 → 0.806 (its turns are the quote-heavy,
 punctuation-heavy ones), kiro_cli 0.202 → 0.520, codex 0.168 → 0.475, grok 0.221 → 0.588.
@@ -76,11 +92,13 @@ punctuation-heavy ones), kiro_cli 0.202 → 0.520, codex 0.168 → 0.475, grok 0
   because the harness still imported the deleted planner — the two were then landed together,
   which is the correct unit).
 - `agent-session-tools`: **1,833 passed** on the final tree (1,805 before Stage 2, +28).
-- `studyloop` package: 3,912 passed; 166 failed / 147 errors, none caused by Stage 2 — 157 are
-  one leaked event loop (`asyncio.run() cannot be called from a running event loop`) that also
-  hits `memory_search`/`get_active_topics`/`log_struggle` tests; the rest are sandbox conditions
-  (no PyPI, no `tmux`, PTY spawn timeouts, journey-root guard). The only file whose failures name
-  `session_search` passes 10/11 in isolation, the one red being `tmux`.
+- `studyloop` package, **matched control**: the pre-Stage-2 tree (`80ee57bb`, worktree, same
+  sandbox) fails 305 node ids; the Stage 2 tree fails 313; **299 in common**, identical exception
+  profiles (158 `RuntimeError` event-loop leaks, 15 `TimeoutError`, 14 journey-root guards, 11
+  `MultiplexerError` = no `tmux`, 11 `CalledProcessError` = no PyPI, in both). The 14 only-in-Stage-2
+  ids are wheel-build smoke tests (network), one live-provider test, and two lifecycle ids that
+  differ only by stderr noise concatenated into the node id; none touch search. Suspected
+  environmental; not demonstrated clean beyond this comparison.
 - Toy corpus (CI validation tier): 7/7 on measured values.
 
 ## Deferred, with reason
@@ -102,6 +120,25 @@ punctuation-heavy ones), kiro_cli 0.202 → 0.520, codex 0.168 → 0.475, grok 0
 - Ranking, not vocabulary and not crashes, is the dominant miss on the corpus the product serves:
   1,832 ranking misses vs 257 vocabulary misses.
 
-## Council addendum
+## Council addendum — single seat `openai.gpt-6-astra`, verdict **REJECT** (6 findings)
 
-_(appended when the single-seat review returns)_
+Brief 25.5 KB; 58 s; 844 words. Every finding verified at source before acting.
+
+| id | sev. | finding | disposition |
+|---|---|---|---|
+| F1 | MAJOR | fallback after a rejected explicit query re-entered explicit detection; `fts:fts:alpha?` runs `alpha?` unguarded | **CONFIRMED at runtime** (reproducer crashed with `fts5: syntax error near "?"`). Fixed: the fallback calls `plan_natural_language()` only, strips every prefix; regression tests through MCP and CLI (`fts:fts:alpha?`, `fts:"alpha" OR ? AND`) |
+| F2 | MAJOR | an uppercase operator inside a quoted span classified as explicit, so `"error OR warning" recovery` never widened | **CONFIRMED.** Operators now count only outside double-quoted spans (`_has_operator_outside_quotes`); tests pin the seat's discriminating case (upper == lower inside quotes; operator outside quotes still explicit) |
+| F3 | MAJOR | census attribution overclaimed from a 400-sample; mechanism unverified | **ACCEPTED** → full-population per-question transition receipt; mechanism read from the frozen escape function (1,217/1,227); zero regressions across 5,434; bound and instrument difference stated |
+| F4 | MAJOR | golden gate labelled met when it was changed; "five added lines" vs 15-line stat | **ACCEPTED** → gate marked NOT MET AS WRITTEN and amended to projected equality, checked between named commits by script; the 15 lines are trailing-comma changes on lines that also gained a comma |
+| F5 | MINOR | 91/91 and "5 rescued" inferred, not counted | **ACCEPTED** → counted directly (`gold_counters`): 91/91 ordered lists, 5 of 42, `moved_outside_crash_class: []` |
+| F6 | MINOR | studyloop attribution lacked a matched control | **ACCEPTED** → pre-Stage-2 worktree run in the same sandbox; 299/305 common ids, identical exception profiles |
+
+After the corrections the code fix and tests are in `c8c105c4`; the gold and census receipts
+were **not** regenerated for F1/F2 because neither reproducer pattern occurs in the gold set or
+in the census population in a way that changes an outcome (checked over all 5,434 texts: zero start
+with `fts:`; **4** carry an uppercase operator inside quotes — all four re-run through the fixed tool
+keep their recorded own-session outcome, 3 hits and 1 miss, now via the `or` plan). The receipts
+therefore stand exactly.
+Escalation rule: a REJECT with MAJORs escalates to three seats — **pending Andy's call** whether
+to run the three-seat council now or fold Stage 2 into the Stage 3 three-seat review already
+scheduled.