@chrono-meta/fh-gate 1.4.97 → 1.4.98

package/knowledge/shared/harness-core/field_verdict_crossfamily_gate.md

@@ -1,9 +1,9 @@
 ---
 name: field_verdict_crossfamily_gate
-description: Load-bearing field-project verdict/gate/safety code gets the same cross-family adversarial gate as FH assets — the correlated default-toward-PASS blind spot is model-family-level, not FH-specific.
+description: Load-bearing field-project verdict/gate/safety code gets the same cross-family adversarial gate as FH assets — the correlated default-toward-PASS blind spot is model-family-level, not FH-specific. §7 adds the standpoint axis (orthogonal to family) for shared-body/cross-harness-boundary changes.
 type: reference
 date: 2026-07-03
-tags: [cross-family, decorrelation, verdict-binding, field-harness, degrade-direction, correlated-blindspot, mode-d]
+tags: [cross-family, decorrelation, verdict-binding, field-harness, degrade-direction, correlated-blindspot, mode-d, standpoint-axis, cross-harness-boundary]
 ---
 
 # Field-Harness Load-Bearing Change Gate — cross-family, pre-merge
@@ -141,3 +141,216 @@ converge loop **before it is considered done**. The autonomy floor applies — t
 judgment on borderline cases is trusted only at opus-tier or above; a below-floor orchestrator
 runs the review or asks, never silently skips (`[[feedback_judge_robustness_mechanical_anchor]]`,
 CLAUDE.md §Floor governance).
+
+## §7 Standpoint axis — orthogonal to family, formalized 2026-08-14
+
+**The gap this closes**: §3 above establishes that cross-family review decorrelates the
+*reviewer's error distribution*. It does not decorrelate the *reviewer's ground-truth source*.
+Three families reviewing a shared-body change (code authored in one harness, consumed by or
+governing interaction with a different harness) all read the SAME artifact from the SAME
+standpoint — the author's own repo, the author's own understanding of the target's rules — if none
+of them ever executes as, or is run by, the target harness itself. A defect that only manifests
+relative to the target's actual conventions is invisible from that standpoint regardless of how
+many families read it, because the miss is not in *how the diff was read*, it is in *what the
+review was grounded against*.
+
+**One sentence**: *family diversity raises resolution within one standpoint; standpoint diversity
+changes which ground truth the review is checked against — they are orthogonal axes, and a review
+that maxes out the first while leaving the second at zero has not raised its coverage of
+standpoint-dependent defects at all.*
+
+**Why the same author cannot close this by working inside the target repo and opening the PR
+there — corrected framing (2026-08-14, after a cross-harness standpoint review of this section
+itself).** The claim is *not* "the author is structurally blind to their own work" — a controlled
+same-repo trial recorded outside this repo found that authors explicitly asked to self-check their
+own output largely close the gap themselves (a clean replication refuted the stronger
+"non-self-administrable" reading). What that trial's surviving result actually supports is
+narrower and still real: a **routinely-run rubric that checks against the target's own corpus**
+does work the author's default self-review does not reliably do on its own — the value is in the
+rubric being supplied and run as a matter of course, not in some inherent authorial blindness.
+This is measured, not asserted, from a second angle too: a 2026-08-14 trial recorded a human
+insider review approving a change that violated a directory convention its own governance
+documents disagreed about — neither reviewer, human or AI, held both conflicting documents at
+once. Standpoint-grounded review is a routine cross-check against the target's ground truth, not
+a claim that the author (of any kind, including a human insider) is incapable of finding it alone.
+
+**Field evidence — three artifacts, two organizations, three repos, one carrying two independent
+trials, one day** (corrected count, 2026-08-14 — an earlier draft of this section said "four
+independent instances"; row 2 and row 4 below are two trials of the *same* artifact, not two
+artifacts, and internal project codenames in a sibling field harness's private tooling are
+genericized here per that harness's own residency rule — the specific names are on record in FH's
+private companion store, not in this shared-layer doctrine file):
+
+| # | Where | What only-standpoint-review caught | What full-family review missed it |
+|---|---|---|---|
+| 1 | forge-harness PR #368 | Write-side namespace fix left the read side (`sync-from-be.sh`, `fh_session_load.sh`) completely unsuffixed — a sibling hub's first SessionStart would read the author hub's own data as its own. **Confound named, not smoothed over**: the file is FH-owned, so "different harness" and "not the original author" are not cleanly separated here — this instance supports *"a different-standpoint check caught it,"* not *"a different-standpoint check was structurally necessary"* (see `fh_three_layer_canon.md §1-b ⓑ` for the fuller hedge) | 2 Claude waves + codex(gpt-5.5), all reviewing from the author hub's own content-reading standpoint |
+| 2 | a sibling field harness's black-box regression repo, PR #8 (reps=3, 2026-08-07) | First measurement of the **finding-class split** under this specific frame — peripheral finding class differs by harness even when core defects are harness-invariant. Not the first sample of the underlying axis: `harness_verification_core_extended.md` (PR #225, 2026-08-01) already doctrinized "verification dispatched from a different harness" as Extended, with its own N=2 evidence; this instance and `[[feedback_decorrelation_axis_matches_failure_mode]]` are additional samples of that same axis, not a new one | Same-model, same-prompt, harness-blind arm |
+| 3 | qasp-dev PR #161 (`mirror_guard_check.sh`) | 67 tracked files sat unregistered in a real protected-path list for months; 50 existing regression fixtures were **synthetic**, built by the same process that wrote the implementation, so they shared its blind spot structurally. **Confound named**: this is a same-author/same-repo blind spot (synthetic fixtures written by the implementer), not a cross-harness-boundary case under this section's own trigger below — an *adjacent* axis to standpoint, not a direct instance of it; kept here as corroboration of the family, not as a fourth standpoint data point | The suite's own author, repeatedly, across its whole synthetic-fixture lifetime |
+| 2b | the same sibling harness's PR #8, known-answer calibration trial (2026-08-14) | 9 confirmed findings beyond a 21-comment human review baseline (7 of them S-tier, verdict-input-corrupting); the trial's own hint-vs-no-hint arms scored identically, so **prompt** variation (the hint) contributed nothing the standpoint shift didn't already carry — the arms did not vary model family, so this instance is silent on the family axis specifically | A target-harness-native human reviewer, plus the same-repo author |
+
+Instance 3 was found *by this session, independently,* while reviewing qasp PR #161 for merge —
+not fed in from the sibling harness's thread — even though (per the confound above) it is adjacent
+rather than a clean standpoint instance. The genuine standpoint pattern (rows 1, 2, 2b) reproduced
+across two organizations the same day it was being formalized, and a fourth, retrospective data
+point landed the same day too: a cross-harness standpoint review of *this very section*, run
+against a sibling field harness's own repo, independently surfaced the residency and citation
+defects this revision fixes — a second live demonstration of the same axis, folded into the fix
+rather than tabulated as a fifth row (it reviewed this doctrine, not a code change). That same
+review also revealed a *sequencing* defect, not just a content one — see **Sequencing** below;
+one event, cited twice for two different things it showed.
+
+**Formalization decision (operator, 2026-08-14)**: the evidence bar this repo requires before
+mechanizing a judgment call (`[[feedback_evidence_threshold_build_discipline]]`,
+`[[feedback_mechanize_at_repetition_prose_before]]`) is met — three artifacts (one with two
+independent trials), cross-organization, with a named causal mechanism, not a recurrence count
+alone. The prior hold on this (pmh-dev#68: *"1회 시행부터, 지금 게이트/마커를 늘리지 않는다"*) is
+**lifted**. This section is that formalization.
+
+**Naming note**: this field's name collides with FH's own existing use of "standpoint" as the
+persona/viewpoint organizing noun (`fh-meta:beginner`/`main-player`/`expert`, "parallax-compatible"
+output — a *different* axis: which persona is looking, not whose repo is the ground truth). Kept
+as-is rather than renamed — the term is already load-bearing in this session's own conversation and
+in cross-harness correspondence about it — but the two senses are genuinely different concepts and
+must not be conflated: agent-registry "standpoint" = which persona reviews; this field's
+`standpoint:` = whose repo the review is grounded against. If this collision causes real confusion
+in practice, the fallback is renaming the *field* (not the section) to `ground_truth:` — the enum
+values below are unaffected either way.
+
+**The field**: `standpoint:` — sits alongside `crossfamily:` in the same load-bearing verification
+marker, never replacing it. Closed enum, same discipline as `crossfamily:`'s three-way *could
+not / did not / did not look* split (a free-prose field would let an unrun check read as a clean
+pass):
+
+```
+tier1                        content-only review — no standpoint decorrelation (the default
+                              unless upgraded; NOT itself a failure, most changes have no target
+                              standpoint to borrow)
+tier2(<target-harness>)      peer-simulated — the reviewer instantiated/ran the TARGET's own repo
+                              (a local clone, real content) and executed the change from that
+                              standpoint. Closes shared-body-path defects; a BARE clone cannot see
+                              the target's gitignored local wiring (settings, consent bindings,
+                              node-local state) — that gap is inherent to a clone, not a defect in
+                              a given run. Named exception, not a loophole: if the reviewer's own
+                              node ALSO mirrors the target's gitignored state through some other
+                              channel (e.g. a companion-store sync that carries `tracks/_meta`
+                              across machines), that visibility is real and should be credited —
+                              but state so explicitly on the marker line, since the default
+                              assumption for an ordinary clone is still "local wiring invisible."
+tier2b(<target-harness>)     same operator, target's real runtime — the SAME human operator who
+                              authored the change runs it in the target harness's actual runtime
+                              (not just a clone's content), so local wiring IS visible, but the
+                              reviewer is not an independent party. Distinct from tier2 (content
+                              only, no local wiring) and from tier3 (independent party). Exists
+                              because tier3 is structurally unreachable for a harness pair with a
+                              single shared operator — see the residual below — and collapsing that
+                              case into tier3 would overclaim independence it does not have.
+tier3(<target-harness>)      actual peer — a DIFFERENT human operator of the target harness ran
+                              the change in their real runtime. The only tier with both local
+                              wiring AND reviewer independence from the author.
+not-applicable                the change has no target-harness standpoint to borrow (no
+                              shared-body surface, no cross-repo consumer contract) — distinct
+                              from a degrade value; this is a scoping fact, not a miss. Carries the
+                              same substantive-grounds-on-the-same-line discipline as a degrade
+                              value below — asserting non-applicability without naming what was
+                              checked is indistinguishable from UNKNOWN wearing a permissive label
+DEGRADED_NO_TARGET_ACCESS     could not — applicable, but no local clone/access to the target
+                              harness existed
+DEGRADED_NOT_RUN              did not — target was accessible, standpoint review was skipped
+UNKNOWN                       did not look — applicability itself was never assessed
+```
+
+**Residual this enum split names rather than hides**: for a harness pair with one shared human
+operator (this repo and a sibling field harness the same operator also runs), `tier3` is either
+unreachable or collapses into "the same author ran it in the other repo" — which is exactly the
+standpoint the "why the same author cannot close this" argument above says is insufficient on its
+own. `tier2b` is the honest reachable rung for that pairing; do not inflate a `tier2b` run to
+`tier3`, and do not undersell it to `tier2` either — it is a distinct, real, if operator-correlated,
+data point.
+
+**Mechanization status — `standpoint:` is prose-only today, and this must not be read as more than
+that.** `crossfamily:`'s degrade triad is hard-blocked at commit (`templates/.git-hooks/pre-commit`,
+`scripts/test_marker_crossfamily_lanes.sh` — grep-verified: ~20 crossfamily references in the hook,
+validated fixtures). `standpoint:` has **zero** matches in that hook and no fixture suite — nothing
+stops an author from writing `not-applicable` with a thin justification, or omitting the field
+entirely, and no marker-shape check catches it. §Marker required fields in
+`.claude/rules/fh_4axis_gate.md` does not yet list `standpoint:` either (add it there when this
+mechanizes). This is the honest current state, not a placeholder apology: the field exists so a
+human reader can ask for it and so the *next* occurrence of a false `not-applicable` has something
+concrete to point at — mechanize on that first recorded false value
+(`[[feedback_mechanize_at_repetition_prose_before]]`), not before. **Ownership**: this field lives
+in FH's own shared-layer canon, so mechanizing it (hook lanes, fixtures) is FH's job — a sibling
+field harness that syncs this file verbatim cannot add the check locally without its own sync
+process rejecting the divergence, so do not expect the check to appear from the consuming side.
+
+**Where the evidence itself lives — the same gate-locality problem, applied to this field.** A
+sibling field harness's own governance doctrine already names this exact defect: verification
+evidence recorded in a gitignored local marker is *"evidence placed where the reviewer cannot
+read it"* — a reviewer on another session, repo, or runtime structurally cannot reach it. This
+field inherits that problem in its sharpest form, because a `tier2`/`tier3` value is a claim
+*about a second party*, and the only party positioned to falsify it is the party the field is
+gitignored away from. Treat the local marker line as a private note, never the canonical evidence:
+the canonical, reviewer-visible copy belongs in the sanitized PR-body evidence capsule
+(`.claude/rules/fh_4axis_gate.md` §Reviewer-visible evidence — same discipline, not a new one), and
+a `tier3`/`tier2b` claim should carry a counter-artifact reachable from the target side (a linked
+issue/PR comment, not just an assertion in the author's own repo) whenever one exists.
+
+**Sequencing — runs on the local diff before the first push, one rule, no risk-branch (added
+2026-08-14, self-correction).** §4 above titles the whole gate *"before merge, not after"* — this
+field **tightens** that, it does not merely inherit it: a push to a public remote is itself a
+publication event, so "before merge" is not early enough on its own (a PR can sit open, reviewed,
+un-merged, and still have leaked). The retrospective review recorded above as the *"fourth,
+retrospective data point"* — the cross-harness standpoint review of this very section — is the same
+event this paragraph is about, cited there for its evidentiary weight, cited here for what it
+revealed about *timing*: it ran on **PR #370**, the PR that introduced this field, **after** the PR
+was already open and pushed to a public repo. That review found four S-tier findings; two were
+residency leaks (an internal codename, a re-identifiable colleague anecdote), and by the time they
+were caught, the leaking lines had already sat in a public, pushed commit — precisely the ordering
+the repo's own Pre-Publish Surface Gate exists to prevent (*"scrub before publish, never
+publish-then-scrub"*). Those specific lines were fixed forward in a later commit, not removed via
+history rewrite (which would itself be a Destructive-Op-gated action) — the pushed commit that
+originally carried them is still reachable in git history. Record which path was taken whenever
+this recurs; do not let "fixed" imply the exposure itself was undone. (Findings recorded in PR
+#370's sanitized evidence capsule; the specific leaked strings are in FH's private companion store,
+not here.)
+
+The fix is not a second risk-judgment ("is this specific change risky enough to justify running
+pre-push instead of post-PR") — that would just add another judged branch point with its own cost
+and its own failure mode. The fix is a single unconditional rule: whenever the §7 trigger below says
+`standpoint:` applies at all, the review runs on the **local diff, before the first push to any
+remote** — public or private, no visibility judgment to make. Axis 2/3 (steel-quench/phantom-quench)
+run earlier still, at first commit (`.claude/rules/fh_4axis_gate.md`); run standpoint alongside them
+when convenient, but the binding line for this field is the push, not the commit. Opening the PR is
+the reviewer hand-off, and by that point this review should already be clean; if it is not clean
+yet, the PR does not open yet.
+
+**Trigger — narrower than §4's full gate, deliberately, and defined by effect, not by file-class.**
+A file-class trigger ("touches `scripts/`, `knowledge/shared/`, `templates/`...") overtriggers for
+a harness pair whose consumer contract already treats nearly the entire shared layer as
+synced-verbatim — for such a pair almost every commit would qualify, which is the over-pricing
+this trigger is trying to avoid, not invoke. The trigger is therefore the **behavioral** subset:
+`standpoint:` is required when a change alters another harness's actual behavior, gate outcome, or
+interaction contract — not merely when it touches a synced path. An ordinary load-bearing change
+with no behavioral cross-harness surface is `not-applicable` by scope, not by degrade, even if the
+file it lives in happens to be synced elsewhere. This mirrors a sibling harness's own scoping
+proposal (pmh-dev issue #68, verified verbatim in that thread: *"대상 후보가 좁다 — 두 허브가 같은
+컴패니언·같은 스크립트를 공유하는 경로. 전면 도입이 아니라 이 클래스만"*) — narrowed here to the
+behavioral reading after that same review found the file-class reading false for at least one real
+pair.
+
+**Relationship to `harness_verification_core_extended.md`'s core/extended axis**: tier2 and tier3
+are both "extended" in that document's sense (they require a cluster member's engine or repo to
+discharge) — this field does not compete with that doctrine, it subdivides one corner of it.
+Extended asks *was a cluster instrument used*; `standpoint:` asks, given that a cross-harness
+surface exists, *whose ground truth was the review checked against*. A verification can be
+Extended (a field harness's own audit lens was dispatched) while still being `standpoint: tier1`
+if that dispatch never executed as the *specific other harness on the other side of this exact
+change* — Extended is about instrument sourcing, `standpoint:` is about whose repo the check ran
+against.
+
+**Named residual, honestly scoped**: the proven-uplift bar `auto-decorrelation` already requires
+for the family axis (`tracks/_meta/decorrelation_calibration_*.md`, N≥3 before claiming *proven*
+rather than *measured*) applies here too — three artifacts (one with two independent trials, plus
+the retrospective standpoint review of this section itself) is enough to mechanize the **field**,
+not enough to claim the standpoint axis's uplift is calibrated in the same statistical sense family
+diversity is. Record accordingly: `standpoint:` entries accumulate toward that bar, they do not
+presuppose it already cleared.

package/knowledge/shared/harness-core/ship_readiness_gate.md

@@ -140,6 +140,23 @@ engine was precisely that **an identity declaration is not a judgment circuit**
 net loss; removing it recovered +0.67 on the weak tier). A one-word translation re-fuses exactly what the
 measurement separated.
 
+**Numbering rule — engines are named, never numbered.** Write `context-continuity`, not "engine ④".
+There is no canonical engine order: the table above reads *external-grounding · judgment-circuit ·
+ship-gate · context-continuity*, and the prose further down this same file reads *judgment-circuit ·
+ship-gate · context-continuity · external-grounding*. A number derived from one decodes to a different
+engine under the other — and the two candidates for "④" are **context-continuity and
+external-grounding**, which currently hold different grades, so the ambiguity is not cosmetic.
+The collision is worse than one file's internal disagreement: **the identities are numbered ①–⑤ and the
+engines are not numbered at all**, so a bare ① in any record is undecidable without its sentence. This
+file shows it — a paragraph enumerating engines sits four lines above `③⑤ are 🟢, ①②④ are 🔵 RC`,
+where those numerals mean *identities*.
+
+> **Legacy decode (do not delete — session records already use numbers).** Records written before
+> 2026-08-13 say 엔진 ①~④. They decode by the **engine table order above**:
+> ① external-grounding(물어보기) · ② judgment-circuit(영혼) · ③ ship-gate(품질게이트) ·
+> ④ context-continuity(맥락유지). That is the ordering those records were written under; it is
+> recorded here so they stay readable, **not** to make the numbering canonical. New writing uses names.
+
 **Why engines gate the advertised capabilities**: the harness's most-advertised surfaces — incubating a new
 project, orchestrating a multi-harness cluster — are simultaneously *long, autonomous, novel and shipping*.
 They therefore load all four engines at once, which is why a harness with a mature ship-gate and little
@@ -151,6 +168,101 @@ else appears to fail *at* those surfaces while the cause is underneath them.
 > what is known. (The counterweight matters here: a mapping that looks tidy is the easiest thing to start
 > citing as a finding.)
 
+### Engine status (2026-08-13) — the first *measured* grades
+
+These grade the **engines themselves**, on the same ladder as the identities (🔵 RC = implemented ∧
+known-pair-calibrated ∧ its own self-test green · 🟢 = a real-situation firing artifact, n≥1). They were
+**measured, not composed**: every leg-2 verdict comes from a **revert probe** — disable the mechanism,
+re-run the suite, check that *exactly* the matching lane reddens — because a green suite is not evidence
+that the suite measures the thing. **What a revert probe proves is lane discrimination** — that a
+given branch has a lane which notices its removal. It does **not** prove that the engine's whole
+instrument surface is covered; those are different claims, and ship-gate below is the row where
+they visibly diverge. That method earned its place the same day: **three of the four engines
+had at least one live branch that survived deletion with every lane still green.**
+
+**Grade = the LOWEST leg that fails, never the highest leg reached.** A real-situation firing does
+not lift a row whose leg 1 or leg 2 is broken. external-grounding below is exactly that case — it
+*has* a genuine firing artifact and is still 🟡, because the layer that supervises the firing sits
+under the RC bar. Reading the ladder as "highest leg wins" inverts every row in this table.
+(Named by the cross-family reviewer as the one real inconsistency in the first draft, which stated
+🟢 = firing n≥1 without saying that the lower legs still gate it.)
+
+⚠️ **These are not identity grades, and they do not upgrade the mapping.** The identity↔engine column
+above remains the unverified hypothesis it declares itself to be. An engine grade says what the harness
+*can do*; it says nothing about which identity that unlocks.
+
+| Engine | Grade | Leg 1 — implemented ∧ wired | Leg 2 — known-pair, revert-probed | Leg 3 — self-test | 🟢 real firing |
+|---|---|---|---|---|---|
+| **ship-gate** (품질게이트) | 🟢 **GREEN** | pre-commit + pre-push, `core.hooksPath` verified live | ✅ 37 cross-family fixtures · 30 branch-claim · marker-floor. **2 revert arms, each reddening only its own lane** (neutering the degrade-grounds check surfaced `'client error'` passing on the substring `cli`; restoring the `single-family` free exit surfaced two more) | ✅ 3 suites `rc=0` | ✅ **twice in one session (2026-08-13)** — a commit blocked as `🚫 BLOCKED — resolve failing axes`, and a branch-claim block that stopped a commit from landing on a **peer session's branch** in a shared checkout |
+| **context-continuity** (맥락유지) | 🔵 **RC** | `compaction_probe` (PreCompact + UserPromptSubmit; snippet ships and `install-wizard` merges it by glob, not by name) · `session_close_check` (pre-push) · `digest_landing_check` · `utterance_landing_check` | ✅ 47 pairs after the 2026-08-13 fix, **4 revert arms**; session axis probed separately (deleting the card-last *verdict* — not its message — reddens its lane) | ✅ 47 · 10 · 8 · 8/8, all `rc=0` | ❌ **withheld, and the reason is the interesting part** — see below |
+| **external-grounding** (물어보기) | 🟡 **PARTIAL** | `novelty_claim_check` wired (pre-commit + selfcheck) but **advisory, non-blocking** · `digest_landing_check` has **zero callers** · the daily digest launcher ships a **placeholder path** in its plist | ⚠️ split: novelty **8/8 arms anchored**; landing-check has **4 live branches that survive deletion**, one of which flips a genuine *miss* into a false *landed* | ✅ novelty 13 pairs · landing 10 lanes — but the latter only runs when a human types it | ✅ **exists** — 5 `frontier-auto:` commits (2026-06-22 → 07-28); one hand-verified, it labelled an unreachable source `UNCALIBRATED` instead of asserting through it |
+| **judgment-circuit** (영혼) | 🔴 | ⚠️ the lint + registry exist and are wired, but the engine is a **6-step loop and only step ③ has a mechanical anchor (1/6)** | ⚠️ known-positive/negative both present; **5 arms unanchored**, including deleting the pre-commit block, the selfcheck wiring, and the registry file — none reddens anything | ✅ `rc=0`, 15 pairs | ❌ not found in this repo's tracked history (controls run) |
+
+**Why ④ is held at RC rather than promoted.** Its three RC legs stand. The 🟢 leg is withheld for two
+reasons that point the same way: (a) **the instrument that would evidence it is disproven** — the scoring
+leg greps the transcript, and the transcript preserves history across a compaction, so that scorer reports
+zero loss forever (fail-open); the isolated scorer that could answer *"can the model still answer?"* is
+unbuilt; and (b) the one observed pre-fix real firing **delivered a false ledger** — a 5-day-old seal
+announced as "the compaction just before this one," which misled the session that then fixed it. After the
+fix the same path prints an honest *"cannot tell"*. Honest inability is not preserved continuity. **An
+engine whose own measuring instrument is refuted cannot be promoted by argument.**
+
+**What the four measurements found in common — one root, three engines.** The wiring of an anchor whose
+lanes live *inside* the script (`--self-test`) is **structurally invisible to the repo's wiring checker**,
+whose scope is the filename patterns `scripts/test_*.sh` / `*_lanes.sh`. So a caller line can be deleted
+and nothing reddens: measured on the judgment-circuit lint (removing its pre-commit block, its selfcheck
+entry, or its registry file each left seven checks green) and structurally true of the landing checker,
+which has no caller at all. This is one rung above the debt the checker was built for: it catches *"a lane
+that never runs"*, not *"a lane that runs, whose caller can vanish unnoticed."* Sharper still — the
+selfcheck comment that closed this class on 2026-08-08 did so with a **hardcoded three-name list**, and the
+landing checker was born the next day outside it. **A repair that enumerates instead of deriving reopens
+itself on the next addition.**
+
+**What blocks the next rung, per engine** — cheapest first, and none of it is a rewrite:
+
+```
+② judgment-circuit  ①②(interview · form-forcing) have ZERO mechanism; ④'s mechanism exists and is
+    🔴 → 🟡        simply not wired to this loop; the registry holds ONE entry (this repo's own
+                    CLAUDE.md), so the instrument has no corpus to measure. Note the instrument
+                    measures the PRESENCE of a declared form, never whether a circuit is real —
+                    its own header says a "default to PASS" direction still earns credit.
+① external-grounding  wire the landing checker (zero callers today), and anchor its four surviving
+    🟡 → 🔵         branches — especially the self-reference filter, whose removal turns a real miss
+                    into a false landing (optimistic direction).
+④ context-continuity  build the isolated scorer. Nothing else moves this row: the question
+    🔵 → 🟢         "was the thread preserved?" has no instrument, and a firing without one is an
+                    anecdote either way.
+③ ship-gate         already 🟢. The open work is not promotion but scope: its own axis-1 job is
+    🟢              still not a required server-side check.
+```
+
+**A 🟢 engine can still hold an instrument with a silent hole — say so rather than letting the grade
+cover it.** This measurement's leg-2 probe for ship-gate covered the marker and branch-claim lanes, not
+every instrument the engine owns. A parallel axis measured, the same day, that one of the others — the
+package-coverage checker — was dropping **every `.json` reference** through an alternation-order bug
+(`js` matching before `json` in a leftmost-first alternation, leaving a path that then fails an existence
+test and is discarded in silence). That checker had never seen a JSON reference in the shipped docs;
+repairing it surfaced three immediately — one of them a **consumer-facing defect**: a shipped
+document instructs the user to copy a settings file that was not in the package at all. That is
+the very class this checker exists for, and its own regex kept it invisible. The engine grade is unchanged — it is earned by blocks that
+actually fired — but *grade* and *instrument coverage* are different claims and must not be read off
+one another.
+
+**And one failure mode this measurement did not anticipate, found the same day by a parallel axis.**
+`npm publish` packs the **working tree, not the commit**. In a checkout shared by several concurrent
+sessions, the session that publishes therefore ships every *other* session's uncommitted draft. It was
+caught here by the pre-publish scan, not by anyone's care — an operator-private token sitting in an
+uncommitted line of a shipped file. The shared-checkout hazard is usually stated as *"my git operation
+moves your working surface"*, which is recoverable; this is the same hazard reaching an **irreversible**
+surface. The fix is not "remove the token" but **publish from a clean tree at a committed state**.
+
+**Named residuals of this measurement.** The registry/launcher state of *other* installs was not measured
+(one machine, one tree). The daily launcher's real installation lives outside any git tree, so its liveness
+is **UNMEASURED, not zero**. The judgment-circuit loop's canonical 6-step definition **does not exist in
+this repository's public knowledge layer** (grep: 0 hits, control positive) — it lives in operator-private
+notes, which is itself part of why five of its six steps have nothing here to anchor. And every arm ran
+`reps=1`; the scripts are deterministic, but the convention is `reps≥3`.
+
 ## FH's own status (2026-07-14) — NOT yet all-green
 
 Engine column added 2026-08-08 (mapping is the unverified hypothesis flagged above; Status column is

package/knowledge/shared/learnings/subagent_invocations_log.yaml

@@ -1942,3 +1942,58 @@
     ★**H1 규율 이행 확인**: innovator 가 벤더 마케팅 출처의 수치 2건(cross-family 검출률 +40~60% · 공유맥락 온보딩 -35~50%)을 SPECULATIVE 로 격리하고 **어느 제안에도 인용하지 않았다**. 대신 «우리 자체 증거가 더 강하다 — 약한 외부 숫자로 강한 내부 실측을 장식하지 마라»를 명시.
     🟥 **codex 발주 비용 관측**: diff + 산문 프롬프트 1회로 충분했다(오전 축의 7회 352KB 대비). 같은 클래스 결함을 같은 채널로 두 번 보내지 않은 것이 차이다.
     ⚠️ **goal-quench 미발화** — 사이드카 3회 미달(2회)이라 카드 K 의 트리거 조건(3회 초과)엔 안 걸렸다. 이건 규율 준수이지 갭이 아니다."
+
+- date: 2026-08-13
+  session: fh-engine4-compaction-guards + qasp 병렬축 (Opus 5 1M)
+  agents_summary: "4 dispatches: general-purpose×1 (qasp §5 계기 보정 — known-pair 전수, 별 체크아웃 qasp-dev) · fh-commons:quench-challenger×1 (Axis 2, compaction_probe diff) · codex gpt-5.5×1 (같은 diff cross-family, Bash 사이드카) · peer-session SendMessage×2 (재출하 축 파일경계 협상 — 디스패치 아님, 계수 제외)"
+  dispatch_count: 3
+  outcome: accepted
+  evidence: "qasp 축 = 68%(미보정) 을 known-pair 로 보정해 **라우팅 정확도 2.9%(35 중 1)** 로 갈랐다. 신호어 21낱말 중 18 DEAD·1 LIVE(「누락」 혼자 −12) → **목록 튜닝 기대이득 ≈0** 을 실증해 무한달리기를 차단. 컨트롤 ⓑ(FN 16건) 없이 ⓐ만 했으면 한쪽만 쟀을 것 — FN 56%·팬텀 6건(도메인 무관 하드코딩 템플릿 발화)이 거기서 나왔다. 자기정정 2건 보고(자기 출처추적 프로브가 known-positive 를 놓친 것 포함). challenger = M5/R4, 그중 4건 채택 — **선행 0 타임스탬프가 `set -u` 아래 훅을 통째로 침묵시킨다**(불변식 파괴)와 **`>43200` 가지 무앵커**(가지 삭제해도 39쌍 초록)가 최상위. codex = ⓒ 미래 시각 음수 지목, 채택. 산출 = FH PR #359 + 동반 저장소의 계기-보정 핸드오프 1건(2026-08-13)."
+  notes: "★**신규 결함 3건이 전부 «내가 그 세션에 방금 쓴 코드»에 있었고 자력 적발 0.** 수리가 신규 결함의 주된 출처라는 기록된 패턴의 n+1 이다. 되돌림 프로브를 넣지 않았으면 그중 둘은 레인이 초록인 채 통과했다.
+    ★**되돌림 arm D 가 내 앵커 하나를 장식으로 지목했다** — `_age=0` 초기화는 `10#` 하에서 도달 불가라 지워도 47쌍 초록이었다. 지우는 대신 **레인 없음을 코드 주석에 명시**했다(있는 척하지 않기).
+    ★**두 계열이 겹친 지점은 하나뿐이다**(미래 시각). codex=산술/이식성 · 같은-계열=레인 커버리지/주장 정확도 — 탈상관 축이 실제로 달랐고, 한쪽만 돌렸으면 절반을 놓쳤다.
+    🟥 **challenger 가 내 Added-Scope 판단의 «근거»를 정확히 반증했다**(결론은 유지): 사후 귀속이 «불가능»한 게 아니라 `rm`→`mv` 한 줄이면 살아나고, 내가 든 기계화 임계 논거는 그 종류에 적용되는 게 아니었다. 결론(별건 분리)은 그대로 두고 근거만 정정했다.
+    ⚠️ **qasp 축의 자기신고 경계**: 손판정에 기계 앵커가 없어 「진짜 기획 문의 1건」은 자평이고, 정답키 대조(조건 ①)는 범위 밖이다. 경합 가설 2개(MTM ③ 부재 vs TC 자체 파손)를 안 갈랐다 — 그쪽이 스스로 적었다."
+
+- date: 2026-08-13
+  session: fh-reship-lane-debt-wiring (Opus 5 1M, 재출하 캠페인 축)
+  agents_summary: "2 dispatches: fh-meta:challenger×1 (Axis 2, DEBT-12 배선 델타 — diff + 주장 C1~C6 동봉) · codex gpt-5.5×1 (같은 델타 cross-family, Bash 사이드카). peer-session SendMessage×8 (소유 선언 6 + 통지 2 — 디스패치 아님, 계수 제외)"
+  dispatch_count: 2
+  outcome: accepted
+  evidence: "두 계열이 **상위 3건에서 독립 수렴**했다 — ⓐ 팬텀 4건(shipped selfcheck.sh 가 미출하 경로를 새로 참조 → package_coverage FAIL) ⓑ `--list-debt` 반쪽 픽스(bash 3.2 빈 배열, rc=1 재현) ⓒ anchor arm 이 `_ships_per_files` rc=2(UNKNOWN)를 SKIP 으로 접음. ★ challenger 는 **Bash 툴 없이 정적 추적만으로** ⓐ를 예측했고 실행이 그대로 재현했다(레포 selfcheck rc=1). ★ challenger 단독: S2 vendored git 트리 3스위트 거짓 FAIL(`git rev-parse --show-toplevel` 이 바깥 레포를 잡음) — **실측 재현 rc=1 → 수리 후 재측정** · M3 DEBT=0 의 분모가 이름 규약 한정(규약 밖 `--self-test` 4종 호출부 0, 전부 shipped) · M4 exit 2/126/127 무라벨. ★ codex 단독: 헤더 수치 11/8 vs 12/9 자기모순 · indirect-branch 바이패스를 **predicate 로 직접 재현**. 채택 = S1·S2·S3·M1·M2·M3·M4·M6·R1 (9건 수리) · R4 는 명명된 잔여 + 픽스처 고정(동작 유지)."
+  notes: "★**주장(C1~C6)을 diff 와 함께 보낸 것이 결정적이었다.** 반환된 판정이 「부분 반증 5 · 반증 1」로 주장별로 갈렸고, 그중 C2(「소비자 거짓 FAIL 없다」)는 **두 방향에서 다르게 틀렸다** — 소비자가 아니라 소스트리가 먼저 빨개졌고(S1), 소비자 쪽은 plain tarball 이 아니라 **vendored git 트리**에서 터졌다(S2). diff 만 보냈으면 「주장이 틀렸다」는 축이 구조적으로 비었을 것이다([[feedback_decorrelation_axis_is_what_you_send]] 의 이득 쪽 얼굴).
+    ★**두 계열의 겹침은 3건, 나머지는 안 겹쳤다.** challenger=실행환경/스코프 · codex=자기모순/predicate 재현. 탈상관 축이 실제로 달랐다.
+    🟥 **자력 적발 0.** 아홉 건 전부 남이 잡았다. 그중 M1 은 **같은 커밋이 20줄 아래에서 설명하는 바로 그 버그**를 세 번째 사이트에 남긴 것이고, S3 는 **그 fail-open 을 막으려고 만든 헬퍼 30줄 아래에서** 12번 재생산한 것이다. 둘 다 반쪽-픽스 전파경계.
+    ★**배선이 가정을 표면화한다** — S2 는 «배선 전엔 안 돌았으니 틀린 루트가 비용 0이었다». 죽은 레인을 살리는 것은 그 레인이 품은 모든 환경 가정을 처음으로 실행에 노출시키는 일이다.
+    ⚠️ **codex 와 내 tarball 실측이 갈렸다**: 내 plain-tarball 런은 `SELFCHECK: PASS`(rc=0), codex 의 런은 `test_ollama_panel_lanes.sh` 환경 실패로 FAIL. 이 델타 소관 아님(선재·환경 의존)이나 **편한 쪽을 고르지 않고 양쪽을 적는다** — flaky/환경의존 레인으로 잔여 등재."
+
+- date: 2026-08-13
+  session: fh-engine-four-grade-table (Opus 5 1M) — 같은 세션 후반, 위 엔트리의 후속
+  agents_summary: "3 dispatches: general-purpose×2 (엔진① external-grounding · 엔진② judgment-circuit 의 RC 3다리 실측 — 각각 동결 워크트리에서 격리) · codex gpt-5.5×1 (등급표 절 본문 cross-family)"
+  dispatch_count: 3
+  outcome: accepted
+  evidence: "엔진①② = 등급 판정의 **입력**을 격리로 뽑았다. 둘 다 동결 워크트리(bdb298e, 과녁 지문 대조 후 시작·clean 복원 확인)에서 돌았고, **되돌림 프로브를 각자 수행**했다 — ①: novelty 8/8 앵커 생존 ↔ digest_landing **4 arm 전부 무앵커**(자기참조 필터 제거가 진짜 미착지를 거짓 착지로 뒤집는 낙관 방향) + 호출부 0개 · ②: 15 arm 중 **5 무앵커**(pre-commit 블록 36줄·selfcheck 배선·등록부 파일을 통째로 지워도 7종 검사 rc=0) + **6단계 루프 중 기계 앵커 1/6** 을 정본 인용과 함께 계수. codex = 사다리 정의의 **실제 비일관 1건** 지목(🟢=실발화 n≥1 인데 ①이 실발화를 갖고도 🟡) → 「등급 = 실패한 가장 낮은 다리」 명시 규칙으로 반영 + 경계 2건. 산출 = PR #363."
+  notes: "★**세 디스패치 전부 내 판정을 바꿨다.** ①②는 내가 「self-test 초록이니 RC 급」으로 읽었을 자리에 무앵커 9건을 들이밀었고, codex 는 내가 못 본 정의 모순을 잡았다. 자력 적발 0.
+    ★**격리가 값을 한 지점이 명확하다** — 두 에이전트가 각각 자기 미측정 항목을 명시 라벨했고(①: launchd 실설치는 트리 밖이라 구조적 측정 불가 · ②: selfcheck 완주 미실행, W2 판정은 소스 근거이지 CI 재현 아님), 그 라벨 덕에 내가 그 칸을 «0»으로 렌더하지 않았다.
+    🟥 **계수 불일치를 정직하게 적는다**: `session_close_check` ④-e 훅은 오늘 **8 dispatch** 를 셌고, 내 두 엔트리 합은 **6**(위 3 + 여기 3)이다. 차 2건을 귀속하지 못했다 — 훅은 총계만 tally 하고 어떤 spawn 인지 남기지 않아 사후 대조가 불가능하다. **미상 2건으로 남긴다.** (재시도·백그라운드 Bash 가 세어졌을 가능성이 유력하나 미검증이고, 그 추정으로 숫자를 맞추는 것이 이 로그가 막으려는 형태다.)"
+
+- date: 2026-08-14
+  session: fh-pmh-sync-collision-fix (Sonnet 5, pmh-dev#68 field-request pivot)
+  agents_summary: "2 dispatches: fh-commons:quench-challenger×2 (Wave 1 isolated attack on scripts/sync-to-be.sh HUB_SUFFIX fix, Wave 2 frozen-tree re-attack pinned to HEAD after Wave-1 remediation) · codex gpt-5.5×1 (auto-decorrelation cross-family sidecar via scripts/sidecar_wait.sh, operator-approved per-run spend gate, transcript captured)"
+  dispatch_count: 3
+  outcome: accepted
+  evidence: "Wave 1 found [S] no serialization between FH/PMH now sharing one companion-store git worktree (a real hole this fix itself opens), [A] git-add-A fallback could stage the other hub's mid-sync files, [A] undifferentiated refuse exit code breaks Stop-hook cooldown, plus named residuals (two-FH-clone collision, fh_session_load.sh unsuffixed reads) — all S/A fixed and empirically verified (real 10-round/20-process concurrent race, git fsck clean). Wave 2 re-attacked the fix itself and found [S] the stale-lock reclaim was ITSELF race-prone (plain rmdir let two processes both reclaim, one stealing the other's live lock), [A] a date-r failure fell back to epoch-0 (fail-open — every lock reads as stale), [A] the tmp-sweep glob was too broad (real risk of silently deleting unrelated mapped-project content forever). codex independently corroborated the date-r fail-open in different words and separately caught that the lock was placed AFTER the mirror phase already touched $BE (moved earlier as a result). Direct 5-way isolated unit test of the atomic-mv reclaim fix: exactly 1 winner, 5/5 runs."
+  notes: "★**Every S/A finding across both rounds was in code I had JUST written that same session — self-review missed all of it, standard pattern for this repo (feedback_repair_is_the_main_defect_source).**
+    ★**Separately, the pre-commit hook itself (mechanical, not a dispatched agent) caught a real confidentiality leak** — my first design hardcoded the literal sibling-hub name into the public script's case-match and HUB_NAME variable; git grep confirmed zero prior occurrences anywhere in this repo, i.e. a genuinely NEW leak. Asked the operator (AskUserQuestion) rather than unilaterally deciding to override or genericize; operator chose genericize. Redesigned to read sibling-hub identity from a local, gitignored key=value config file instead of a hardcoded literal — and in TESTING that redesign, self-caught a second bug: the new helper function could abort the whole script under set -e/pipefail on the 'no local file' path, never reaching the intended exit-10 refuse message. Neither of these two defects came from any dispatched reviewer — both were self-caught by actually EXECUTING the known-pair suite after each change, not by reading the code.
+    ⚠️ **Live production verification is a named, deliberate residual, not an oversight**: a real .fh-hub-identity.local file was written for this operator's own pmh-dev clone, but the first real cross-hub sync into the shared private companion store was NOT run this session (would be the first-ever real push of PMH content into a store this operator relies on for other tooling) — left explicit for the operator or a future session, not silently done."
+
+- date: 2026-08-15
+  session: fh-mbp-selfcheck-wiring (Sonnet 5, MacBook Pro machine, embedded --self-test DEBT wiring)
+  agents_summary: "3 substantive dispatches: fh-commons:quench-challenger×1 (Wave 1 isolated attack on scripts/selfcheck.sh wiring diff) · fh-meta:challenger×1 (Axis 3 phantom-quench-style grounding audit, same diff) · codex gpt-5.5-xhigh×1 (auto-decorrelation cross-family re-check, live tool execution via scripts/sidecar_wait.sh). Plus 3 non-review codex diagnostic pings (model-identity probes, resolving an operator-flagged model-pin question — excluded from dispatch_count, same as this log's own SendMessage-exclusion precedent)."
+  dispatch_count: 3
+  outcome: accepted
+  evidence: "Wave 1 found 1A+6B on a small diff wiring 4 undeclared embedded --self-test subjects into selfcheck.sh (feeds npm test + prepublishOnly). 5 of 6 B's + the A fixed same round: A(capability_registry_check.sh's calibration marker only appeared in a lane TITLE, not a terminal verdict line — a future rename would flip a real PASS to a false FAIL) · B(rc=10 harness-error mislabeled as 'dispatcher missing?') · B(missing timeout on the loop's largest suite — 45+ lanes, no time bound) · B(exec-failure on a missing sibling file mislabeled as a lane failure) · B(PASS-line rendering could leak cleanup-trap stderr). codex R2 (live tool execution against the post-fix diff, not read-derived) independently re-verified the same 4 subjects' real --self-test output and control/exit-code logic, converged with 0 new findings. A separate mechanical lint (test_selfcheck_state_lanes.sh, not a dispatched agent) then caught a regression the fixes themselves introduced — 2 new raw `tail` blocks bypassing this file's own `_show_failure` helper — fixed and re-verified 44/44."
+  notes: "★**Self-caught 0, again** — the mechanical lint (not either dispatched reviewer) found the regression the S/A fixes themselves introduced, same shape as the pattern this log already tracks (feedback_repair_is_the_main_defect_source).
+    ★**The phantom-quench audit flagged a claim PHANTOM by reading a stale illustrative comment as live ground truth** — the checker's own header names an old example subject (directional_diff_gate.sh) instead of the real live-computed one (relay_channel.sh). Re-derived ground truth by RUNNING the checker's own Python detection logic directly rather than trusting either the audit's reading or the comment: confirmed the diff's own claim was TRUE, but the investigation surfaced a genuinely separate, real bug — lane_runner_check.sh's self-test detector scans a subject's OWN file among its 'runners' search and misreads directional_diff_gate.sh's usage-example comment as a real caller (self-referential false positive), so that subject silently reads as wired when nothing actually calls it. Recorded, not fixed this session (different file/job) — tracks/_meta/fh_signal_2026-08-15_selfcheck-wiring.md.
+    ⚠️ **A same-session self-correction on the codex dispatch itself**: an operator-preferred lighter codex tier (gpt-5.6-luna) was believed to have silently fallen back to a bare 'gpt-5' based on a self-report identity probe ('what model are you'). The operator flagged this as surprising given the CLI's own model picker showed it as a valid current tier. A live control (pinning a clearly-different, smaller model and repeating the same probe) got the identical 'gpt-5' self-report, while the tool's own `model:` banner line was correct every time — proving the self-report probe itself is not a discriminating identity signal for this CLI, not that the pin failed. The R2 dispatch had already redirected to gpt-5.5-xhigh before this was caught, so the cross-family result stands (a clean pass on a stronger model is not weakened by this), but the diagnosis that motivated the redirect was wrong and is corrected in the Axis-2 marker.
+    🟥 **Tally discrepancy, named not forced**: the mechanical SubagentStop tally counted 16 dispatches machine-wide for 2026-08-15; this entry accounts for 3 substantive + 3 diagnostic = 6. The remaining ~10 are unattributed — a live peer session shared this exact checkout for most of the day (confirmed via ListAgents/SendMessage during this session's own branch-claim conflict) and very likely contributed some of them, but the tally has no per-spawn attribution (same structural gap this log already named on 2026-08-13), so the split is not reconstructable after the fact. Recorded as unknown rather than estimated to fit."

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chrono-meta/fh-gate",
-  "version": "1.4.97",
+  "version": "1.4.98",
   "description": "FH runtime adapters — run FH governance, skills, and agents via Claude or Codex with machine-parseable gates.",
   "license": "MIT",
   "keywords": [
@@ -25,6 +25,7 @@
   },
   "scripts": {
     "prepare": "chmod +x bin/fh-gate.js bin/fh-run.js bin/fh-goal.js bin/fh-codex-doctor.js scripts/fh-gate.sh scripts/fh-run.sh scripts/fh-goal.sh",
+    "postinstall": "node scripts/postinstall_notice.js",
     "test": "bash scripts/selfcheck.sh",
     "prepublishOnly": "bash scripts/version_lockstep_check.sh && bash scripts/selfcheck.sh && bash scripts/package_coverage_check.sh --vs-tarball && bash scripts/public_surface_scan_files.sh",
     "release": "bash scripts/public_surface_scan_files.sh && npm publish"
@@ -60,12 +61,15 @@
     "docs/codex-compat.md",
     "docs/pillars.svg",
     "scripts/fh-gate.sh",
+    "scripts/frontier_digest_autopilot.sh",
+    "scripts/postinstall_notice.js",
     "scripts/fh-run.sh",
     "scripts/fh-goal.sh",
     "scripts/count_check.sh",
     "scripts/selfcheck.sh",
     "scripts/version_lockstep_check.sh",
     "scripts/test_selfcheck_state_lanes.sh",
+    "scripts/test_lane_runner_lanes.sh",
     "scripts/test_version_lockstep_lanes.sh",
     "scripts/package_coverage_check.sh",
     "scripts/lane_runner_check.sh",

package/plugins/fh-commons/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "fh-commons",
-  "version": "1.4.97",
+  "version": "1.4.98",
   "engines": {
     "claudeCode": ">=1.0.0"
   },