test(okf): add running n-way per-file parity gate + adversarial corpus (STEG 3)

Make OKF §3 checker parity a RUNNING per-file red/green signal instead of a
stale one-shot claim. The catalog and okr checkers provably diverge on the same
input (catalog walks everything + no BOM/CRLF norm; okr skips innboks/dot-dirs +
normalizes); this gate turns that divergence into a monitored, tested fact.

- scripts/check-okf-parity.mjs: n-way runner, zero-dep. Normalizes each impl's
  existing checkBundle(root) return to {conceptCount, untyped[], okfVersion} and
  compares per file over default read-mode. Impl registry is availability-guarded
  so a catalog-only checkout degrades gracefully; a documented subprocess seam
  awaits llm-ingestion-okf's Python checker (fase 2) + Node port (fase 4).
- test/okf-parity-corpus/: committed byte-exact adversarial corpus + manifest.
  Red-proof fixtures (expected: diverge) the gate MUST go red on — byte axis
  (BOM+CRLF) and tree axis (innboks, dot-dir) each carry one; canon = both.
  .gitattributes -text pins the BOM/CRLF bytes. Green fixtures pin agreement
  incl. a SHARED gap (both descend into node_modules).
- check-okf-parity.test.mjs: 8/8. Corpus assertions + byte-tracking red-proof +
  runtime-materialized symlink/NFC-NFD meta-tests + graceful-skip seam.

Honest scope: only 2 impls expose a runnable per-file checker today, so the gate
runs 2-way and is architected n-way — no faked third impl, no cross-repo code.
spec §7 + log.md updated (no self-certified parity; fixes stay upstream-first).

Verified: parity 8/8, CLI exit 0; regression 33/33; check-versions 10 OK.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NeK9hkxrU9wFPBYGYnSV1V

docs/okf-second-brain/log.md

@@ -161,6 +161,31 @@ protocol; 🟢 is reserved for the independent gate-verified step (operator veri
   today, so okr's migration (move the value to `okf_layout`) is cheap when it happens, and it should
   happen before that tightening. Additive + disambiguating: the §3 floor is unchanged and `okf_layout` is
   optional, so **the convention version stays `0.1`** — no re-ratification triggered. _(catalog session.)_
+- **2026-07-23** — **STEG 3: the running parity gate landed** (`scripts/check-okf-parity.mjs` +
+  `test/okf-parity-corpus/` + `check-okf-parity.test.mjs`). Parity is now a *running* per-file red/green
+  signal over a committed spec-adversarial corpus, not a stale one-shot claim (retracted above). **Verified
+  scope blocker (ground-truth this session, 2 independent agents + grep):** the round named ≥3 impls, but
+  **only two expose a runnable per-file checker** — catalog + okr. llm-ingestion-okf's Python lib is a
+  *producer* (no `check_bundle`, no CLI; checker is fase-2, its own repo); linkedin-studio's TS §3 logic
+  lives only in private test helpers (not a participant). **Decision (operator delegated on quality
+  grounds):** build the gate **n-way-capable, wire the 2 real impls now**, with a documented availability-
+  guarded seam for the Python checker (fase 2) and its Node port (fase 4) — no faked third impl, no
+  cross-repo code. Decisions the earlier record left open, now fixed: **(a) per-file contract** = each
+  impl's existing `checkBundle(root)` return normalized to `{ conceptCount, untyped[], okfVersion }`,
+  compared per file (concept-set + type verdict), over **default read-mode** (strictIngest off, not
+  files-scoped); **(b) adapter convention** = JS impls imported directly (both export `checkBundle`), zero
+  subprocess/zero dep; cross-repo/other-language impls get a subprocess `--json` adapter when they exist;
+  **(c) corpus form** = committed byte-exact fixtures (`.gitattributes -text` pins BOM/CRLF), the inspectable
+  "borne proof"; **(d) red-proof** = `expected: "diverge"` fixtures the gate MUST go red on (byte + tree
+  axes each carry one; canon = BOM+CRLF concept + `innboks/` raw file, catalog FAIL vs okr OK). Axis
+  decisions: IN = byte (BOM/CRLF), tree (innboks, dot-dir, node_modules[shared gap, both descend],
+  duplicate-name, hierarchical); **WITHDRAWN** = verbatim/trailing-newline (trinn E §1); **OUT of STEG 3** =
+  path-forms (rerouted to §1.2/STEG 5). The git/FS-hostile axes (live symlink, NFC/NFD filename) are
+  covered as **runtime-materialized meta-tests** (temp dir), not committed fixtures. The gate does **not**
+  claim the two checkers are identical (they provably are not) — it makes their divergence a monitored
+  fact. Verified: parity 8/8, `node check-okf-parity.mjs` → exit 0 ("parity holds"); regression 33/33;
+  `check-versions` 10 OK. Fixes stay **upstream-first** (patching BOM-norm into the catalog copy would just
+  re-diverge in the other direction). _(catalog session.)_
 
 ## Distilled architecture notes (OKF round, 2026-07)