P28-02-EQ: 意味を保つ軽量化：同等性の先行固定（harness）

docs/plans/original-completion-2026-10-02/execution/p28-02/eq/PROGRESS.md

@@ -1,4 +1,78 @@
-# P28-02-EQ progress — S5 (repair 10, Claude Opus 5.5 Desktop implementation)
+# P28-02-EQ progress — S6 (repair 12, Claude Opus 5.5 Desktop implementation)
+
+Implementer: Claude Opus 5.5 in the Desktop control session, by the user's delegation
+("Codexの使用量上限いきそうなので、作業をひきとって"; Root relayed 「実装担当含めOpusに委任」).
+The registered contract keeps worker codex; no Codex/Astra run was started for this stage.
+Submission goes through the normal Runner submit, which makes the commit. 正式受入は保留、原本加算0。
+
+- S a46418d064ebf8c8e03043bae50c29ac174b5472 (repair 11): its exact-S native passed
+  (native-a46418d-r01, 7 phases, cleanup errors empty, sealed), and its formal validation
+  cb059f305dd34e7aa811b40ad1cfb6f2 passed both checks (unit 162 tests OK; native-verify
+  exit 0, verified). The fresh independent Opus review (session
+  local_179020e9-e4c8-45ab-aaa0-be27aaa3f3d6, DESKTOP_RESULT.json
+  c08eedb2ff81fd308e0771aeb4330076389561255a2bb964567be28f478592cd) returned
+  changes_requested: EQ1-EQ4 and EQ6 met, EQ5 not met. The normal intake recorded
+  review_failed and repair_pending; the normal repair made repairs 12. All of it is retained.
+- The high finding is accepted: the source closure was only unit-tested. run_eq and verify_eq
+  required the before SOURCE_SHA, so the frozen EQ could not admit an OPT S that changes only
+  registered allowed files (OPT_CONTRACT_T1 §2 and §4). S6 (EQ_S6_AFTER_VERSION_PLAN):
+  - BEFORE_SOURCE_MAP.json holds the registered base product source map (1407 rows); its
+    digest must equal the existing SOURCE_SHA on every read. compare_versions.admit_source
+    admits the before version (no registration, no difference) or an after version (the
+    changed set equals the sorted registered OPT files; no added or missing source).
+  - run_eq --opt-changed PATH (repeatable) registers the OPT files; actual_pins records
+    registered_changes and source_closure, the admission records version and
+    measured_source (before S_eq-native-not-Q, after S_opt-native-not-Q; never Q).
+    verify_eq re-runs the closure with the recorded registration and requires the recorded
+    closure, version and label; actual_pins on the checkout gets the same registration.
+  - verify_eq --compare-versions --before <bundle> --after <bundle> is the bundle-level entry
+    of the comparator (SPEC §2): both seals, versions and closures again, every non-product
+    pin equal, compare_reports on the sealed A-phase reports of R and O, existing defects of
+    both versions listed (after_all_e_pass for the OPT judgment).
+- The three low findings are fixed: the probe records the actual _latest_digest_dirty before
+  each E5 operation (false or null accepted, true rejected); bundled_inputs requires the
+  collector's fixed manifest SHA and normalized absolute destinations inside the read-only
+  directory; the seal (bundle schema v2) records every directory, empty ones included.
+- Evidence kinds are kept apart:
+  - Unit, engine-free, real code paths: admit_source; run_eq.actual_pins with a synthetic map
+    (only collector.source_map, inputs, host SHAs of the synthetic fixture/engine and the
+    machine query are patched); recompute of a real synthetic after bundle; compare_bundles
+    unmocked on a real before and a real after synthetic bundle (positive, a real state
+    difference, pin drift, closure and label tampering, seal tampering, swapped, same bundle).
+    One unit test (S6Admission main) mocks actual_pins only to see the registration main
+    passes; it is not counted as admission evidence.
+  - Diagnostic, not formal (EQ_S6_AFTER_ADMISSION_DIAG_R01 in the Root control folder): the
+    real run_eq.py --check, unmocked, on a throwaway after-version checkout (an APFS clone of
+    this working tree as its own temporary git repository, one allowed OPT file changed and
+    committed). Before version and the registered after version were admitted; no
+    registration, a wrong registration, a registration without change, an added source and a
+    non-OPT change were rejected with the source closure error. No engine was started.
+- Unit evidence (engine-free `python3 -B test_eq.py`; full raw outputs with SHA files):
+  - Class-targeted runs during development were displayed through tail and are not claimed
+    as retained raw results. One of them ran out of disk space (the first S6 bundle tests
+    copied 709 MB synthetic bundles); the tests now clone with cp -c and build two bundles
+    per class.
+  - UNIT_OUTPUT_S6_ATTEMPT_01 (final, all changes included): 179 tests OK in 207.351 s,
+    Python exit 0; same bytes as UNIT_OUTPUT.txt,
+    SHA-256 9dd93de5fd3b66791d6902d51a055caa2c3dd559de6c42019ebb1ab147b2af11.
+  - UNIT_OUTPUT_S6_SEATBELT_01: the same suite under the Runner validation seatbelt profile,
+    run locally (see its exit line in IMPLEMENTATION.md). Local runs are not formal validation.
+- Test bodies changed for S6, each recorded here: the E5 fixture detail carries
+  latest_digest_dirty_before false instead of digest_materialized_before true, and the
+  materialization test rejects true; s4_bundle_fixture takes an optional registration and
+  pin overrides and records commit, version, measured_source, registered_changes,
+  source_closure and the non-product pins; the S4 probe-admission mock returns a before
+  source_closure; the stored-result test records registered_changes [] in its admission and
+  mock; the bundled-inputs test holds its synthetic manifest to its own SHA in place of the
+  collector's and expects the manifest error before the escape error. All other earlier test
+  bodies and raw outputs are unchanged; 17 S6 tests were added.
+
+Next: Runner submit → exact-S native 7 phases (before version, no --opt-changed) → bundle →
+p28-02-eq-unit and p28-02-eq-native-verify → fresh independent Opus review of all 6
+conditions → normal qualification, approval, integration and final → OPT/whole-unit.
+23/88、4/30は不変。
+
+## Retained S5 progress (unchanged historical text)
 
 Implementer: Claude Opus 5.5 in the Desktop control session, by the user's delegation
 ("Codexの使用量上限いきそうなので、作業をひきとって"; Root relayed 「実装担当含めOpusに委任」).