docs(mmp): S4 synthesis + done.md — two-axis parity checklist, root-cause map, prioritised fixes (closes pilot)

[artyhoo]

Summary

S4 synthesis closing the multi-model-pipeline-pilot umbrella: the binding two-axis parity checklist + root-cause map, merging the S2 in-container evidence (merged in #1111), this session's S3 host-side observations, and the §3b banked pipeline-leg rows.

One aif task (e43ce4ea, profile Z.AI GLM-5.2 SDK, off-peak, 552K tokens) carried the whole audit self-referentially — its work product WAS the evidence, and its journey through /arch → /pipeline → /dispatcher → aif/GLM → shields → harvest → acceptance tested every leg.

Also lands done.md (umbrella-closure convention) and a prioritised fix list (§8) — surfaced, not autonomously opened as PRs.

Rollup: 16 ✅ SAME · 4 ⚠️ WORSE · 0 ❌ BROKEN · 18 ◻️ COVERAGE-LIMITED (each with a named would-reach-it carrier). The BROKEN count moved from 3 → 0 by the §5 Q2 analysis, with an explicit reclassification note: inapplicability-to-this-carrier is not breakage, and the ⚠️ WORSE rows (the real defects) are untouched.

Headline finding — check-doc-authority.sh is registered in the container yet neither checks nor complains: jq is MISSING so it skips by its own code, and its warning goes to bare stderr which the harness never surfaces to the agent. A gate that looks alive in any settings audit while being functionally dead — #warning-nobody-reads in its purest observed form. Fix-pointer: install jq in the image AND re-emit via hookSpecificOutput.additionalContext.

Second finding (host-side, new): harvest-via-api.sh mints a fresh commit SHA via the Git Data API, so a cold fidelity audit anchored on the container commit can never satisfy the pr-body-fidelity Audited-SHA == PR head guard. Resolved here by re-anchoring on a byte-identical blob (spec-D10 idempotence); fix-pointer recorded.

Parked questions — all three closed: Q1 RESOLVED (root-caused above) · Q3 RESOLVED (plan ran on GLM, not Opus — the D1 whole-pipeline-on-executor intent holds) · Q2 RESOLVED BY ANALYSIS — re-framed from «what was intended» (human-only) to «what should be true for aif to work well» (answerable from what each hook does): 3 of the 5 unregistered hooks are correctly absent for an autonomous container worker — inject-output-language would be actively harmful if registered, since it pushes operator-facing language onto a worker whose repo artifacts must be English — 1 is low-value, and 1 (inject-subagent-context) becomes a real gap only once workers spawn subagents. «Drift» is the wrong frame; the actionable container defect is the missing dependencies (§8 P0).

Top actionable (§8 P0): install jq, gh, python3 in the aif container image — one root cause behind check-doc-authority skipping its own check and both pre-commit JSON/YAML validators silently skipping. A registered-but-dependency-less gate is the worst failure mode in this codebase's model: alive in every settings audit, enforcing nothing.

Fidelity verdict

FIDELITY: skipped — host-authored S4 synthesis, not an aif-produced stage diff; no ## Provenance substrate is declared, so there is no container diff to audit for WHAT-conformance against a worker kickoff scope. The audited artifact of this umbrella was the S2 evidence, which carried its own cold GO verdict in #1111.

§1.7 Forward-check applied

Checked the disciplines this synthesis operates under:

• no-paid-llm-in-ci.md §1 — all observation was session/aif-bound; nothing audit-related enters CI, and the synthesis is plain markdown. file:line evidence: .claude/rules/no-paid-llm-in-ci.md:1
• ai-laziness-traps.md §2 T14 — reachable-but-unfired rows are recorded COVERAGE-LIMITED with a named would-reach-it carrier, never upgraded to SAME. file:line evidence: docs/meta-factory/research-patches/2026-07-23-aif-parity-s4-synthesis.md:1
• ai-laziness-traps.md §2 T15 — §6 records that the auditor reproduced the very failure class under study (a jq-absent probe whose fallback printed a conclusion), caught only by chasing a contradiction with worker evidence. file:line evidence: docs/meta-factory/research-patches/2026-07-23-aif-parity-s2-container-evidence.md:22
• attention-is-not-a-mechanism.md §1 — the headline finding is classified against #warning-nobody-reads; detection in this run came from a named cold agent + a fail-closed gate, not from anyone remembering to look. file:line evidence: .claude/rules/attention-is-not-a-mechanism.md:1
• doc-authority-hierarchy.md §5 — the patch inherits research-patches folder authority and carries a principle-10 scope marker on line 1.

§1.7 Backward-check applied

Swept sibling surfaces of class «parity census / audit-evidence artefacts» before landing:

• .claude/rules/zcode-parity-doctrine.md — used strictly as a row-set source (T-AIP-D); this synthesis adds the CC-vs-aif axis and does NOT supersede or amend the CC-vs-ZCode census verdicts. file:line evidence: .claude/rules/zcode-parity-doctrine.md:1
• docs/meta-factory/research-patches/2026-07-23-aif-parity-s2-container-evidence.md (merged docs(mmp): S2 in-container parity evidence — aif/GLM-5.2 SDK carrier run #1111) — extended, not superseded: §2 here merges its row-set with host observations; its own §7 worker-checklist is the direct input. file:line evidence: docs/meta-factory/research-patches/2026-07-23-aif-parity-s2-container-evidence.md:235
• .claude/orchestrator-prompts/multi-model-pipeline-pilot/kickoff.md §3b banked rows — carried forward verbatim into §2.2 with their original verdicts; the D1 marker row is updated pre-fix→post-fix rather than rewritten. file:line evidence: .claude/orchestrator-prompts/multi-model-pipeline-pilot/kickoff.md:64
• Acceptance-contour artefacts (feat(acceptance): fail-closed fidelity gate — cold WHAT-conformance auditor + PR-body CI gate (acceptance contour Phase A) #1102/feat(acceptance): choreography wiring — fidelity verdict at every PR boundary + Tier-2 routing amendment (Phase B) #1106) — this run supplies their first live RED+GREEN evidence on a real PR; no contour doc is contradicted, one interaction defect is newly reported (harvest-via-api × Audited-SHA). file:line evidence: agents/fidelity-auditor.md:74
• done.md closure convention — written in this PR per the umbrella-closure convention. file:line evidence: CLAUDE.md:1