arch-v2 S-E: L1 budget gate + config-assertion asserts + InstructionsLoaded verification (fidelity GO — round 4)

[artyhoo]

DRAFT — do not merge. Both cold fidelity rounds returned STOP and the cap is 2 rounds. Round 1's four load-bearing findings are fixed and re-audited CLOSED. Round 2 raised two more, and both are now resolved — but neither by an executor talking its way to GO. picomatch in the shipped hook was proven by CI (a pnpm consumer could not git push at all), fixed, and has since dissolved: removing P2b takes picomatch out of the hook entirely. BLOCKED-1 went to the maintainer, who decided remove P2b after the overlay semantics were settled against the shipped client's source — the client MERGES, so the gate could never fire. fidelity-verdict-in-pr-body is RED on purpose; see the verdict section for why no token is present.

Prior-art: prior-art-evaluations.md#238 (picomatch — ADOPT verbatim; it IS the glob matcher the shipped Claude Code client bundles for claudeMdExcludes path-scope matching, so the problem class is an exact match and no wrapper layer is warranted. The commit pins picomatch@^4.0.4 in packages/core devDependencies — promoting the already-present transitive to a declared dep so an upstream bump cannot silently swap the matcher and change isMatch semantics under principle 34. Row authored in the same commit as the capability artefact, per CLAUDE.md; renumbered 237 → 238 during egress after a concurrently-merged PR claimed 237.)

Harvested from aif task 8068bb5b (status=done 23:06:32Z) via egress Channel A — bundle out of the container worktree, rebase onto live staging, host git push so the real .husky/pre-push ran. It passed, including S-E's own new budget gate (OK: always-on 48671B within ceiling 54000B), and bash scripts/host-verify.sh ran the kickoff's declared contract 3/3 on the host.

Two egress artefacts, attributed so they are not mistaken for stage work:

• SSOT 237 → 238. A concurrently-merged PR took row 237 while this branch also authored 237. The merged row keeps its id; this branch's picomatch row is renumbered to 238, with the Prior-art: trailer and the in-code reference updated to match. Verified: rows 235-238 sequential, exactly one line added, row 237 untouched.
• synth-and-wire.bundle.mjs — regenerated, then REVERTED, and the earlier claim here was wrong. An intermediate commit regenerated the bundle and this bullet claimed the drift was pre-existing on staging. CI disproved it: the regenerated artifact failed the drift gate on this PR while the same gate is green on staging. Root cause was self-inflicted — running npm --prefix packages/core ci inside this worktree replaced the canonical packages/core/node_modules symlink with a real directory carrying a nested semver@7.8.1, and the bundle entry resolves semver from there. Restoring the symlink via scripts/worktree-node-modules.sh made the local build reproduce the committed bundle again (✓ in sync). Reverted in 9277364f6f; the bundle in this PR is now byte-identical to staging's. The verification that was missing: whether a lockfile-faithful install reproduces the committed artifact — not merely whether the top-level version matches the lock.

Round-1 findings — fixed (commit 0eabadc235)

Each was independently reproduced before being fixed; each fix is shown by an observed failure, not asserted.

Finding	Fix	Shown by
BLOCKER-1 relative paths	principle 34 enumerates absolute paths from git ls-files; ships N34-1a, the kickoff's mandated FAIL leg — the live committed list rewritten to the historical relative form must be dead for every entry while the live form stays matched on the same population	reverting the enumerator to relative paths turns N34-1a RED with Live entries: 7; detected dead: 0; restoring turns it GREEN
MAJOR-1 CI mirror not required	alwayson-budget added to ci-success.needs with the reason in-line	branch protection requires only ci-success, so a red budget gate now blocks the merge
MAJOR-2 walkers descend into nested worktrees	both walkers switched to the git-tracked list	measured on the primary checkout: 293,450 files / 8,927 ms → 2,158 files / 12 ms, and a deletion can no longer be masked by a stale copy in another worktree
MAJOR-3 missing doctor arm	scripts/worktree-doctor.sh gained the P2b sweep; it invokes the section through its existing PREPUSH_ONLY=local-claudemd-shadow seam rather than reimplementing it, runs each worktree's own pre-push.ts (REPO_ROOT is script-location-derived), and announces an unrunnable worktree loudly instead of counting it clean	bash -n clean; arm exercised locally
MINOR-3 wrong trigger comment	corrected	—

P2b's discrimination pair — previously INCONCLUSIVE, now closed. A strict-SUBSET local list exits 1 naming the fallen-through file (.claude/rules/cold-seat-economy.md); the same list as a superset exits 0. The operator's gitignored .claude/settings.local.json was restored to its original state (no claudeMdExcludes key) after the fixtures.

Suite after the fixes: 37 files, 352 passed / 1 skipped. host-verify 3/3 on the host. Pre-push green on push.

What changed after round 2 — P2b removed (commit e94be1e5f9)

Round 2's BLOCKED-1 said the REPLACE-PER-KEY overlay verdict rested on a mis-attributed quote
and that kickoff §1 item 4's PARK was not honoured. That is now settled at the source, and the
maintainer's decision (2026-08-07) is remove.

The client MERGES array settings across settings files. Read from the installed binary
(claude.exe v2.1.207, commit bc512d563325; strings -n 8 + grep, every symbol resolved):

_1 = ["userSettings","projectSettings","localSettings","flagSettings","policySettings"]
Soi(): n = eie(n, source, ipe)                  // lodash-mergeWith shape, ipe = customizer
function ipe(e,t,r){ if(Array.isArray(e)&&Array.isArray(t)){ if(r==="fallbackModel") return t; return WSm(e,t) } return }
function WSm(e,t){ return Mo([...e,...t]) }
function Mo(e){ return [...new Set(e)] }

fallbackModel is the sole replace key — exactly what settings.md:278 documents. Scope
precedence orders the fold (local wins for scalars); it does not make arrays replace. So the
effective list is project ∪ local, a local list can only ADD excludes, and P2b's superset
assert held by construction — it could never go red.

Removed: the local-claudemd-shadow pre-push section, its registry entry, its
PREPUSH_ONLY arm, the AIF_CLAUDEMD_LOCAL_SHADOW_ALLOW escape token, and the
worktree-doctor.sh sweep arm. measure-always-on.sh now applies the union, deduped.

Kept: principle 34 (P2a), the packages/core picomatch@^4.0.4 pin and SSOT #238 —
principle 34 imports picomatch directly and never depended on P2b.

Docs corrected in place, not deleted (T18): the overlay verdict patch now records MERGE
with its evidence and falsifier; spec §1.6 FORK D records that its rejected project ∪ local
draft was right; the S-E kickoff records the PARK against the preserved original task text;
the InstructionsLoaded patch drops its stale P2b cross-reference.

The root defect, worth more than the removal. The correct answer was reached twice and
discarded twice. An earlier session concluded "arrays concatenate" and the verdict patch
overrode it as a misreading; a spec draft wrote project ∪ local and spec §1.6 FORK D round-4
MAJOR-3 overrode it because union "contradicts P2b" and "under union it would be vacuous". Both
times the disproof was written down and then used as grounds to reject the model, because the
gate was held fixed and the world model was bent to keep it load-bearing. The mis-attributed
docs quote is downstream of that. Shape to watch for: "model X would make our gate pointless,
therefore not X."

Verification of the removal (host)

Check	Result
npm --prefix packages/core run test:principles	37 files, 352 passed / 1 skipped — identical to the pre-removal baseline
npx tsc --noEmit	clean
bash -n on both edited scripts	clean
host-verify.sh (kickoff's declared contract)	3/3 on Darwin
budget gate	OK: always-on 48671B within ceiling 54000B
host git push	real .husky/pre-push ran green

Meter fixtures — run against a backed-up, md5-verified-restored local settings file:

Fixture	Observed
local adds a new pattern	overlay_source=project+local excludes_applied=8 (union)
local duplicates a project pattern	project+local excludes_applied=7 (dedupe; key still reported)
local is a strict SUBSET — the case P2b guarded	resident set and 48671 B unchanged; no project exclude lost

The subset fixture demonstrates the meter's behaviour. The claim about the client rests
on the source reading above, not on it — claude doctor does not print effective settings, so
no live client-level A/B discriminator was obtained. The earlier "4 local globs work, 3 project
relatives do not" observation is non-discriminating: merge and replace predict identical output
when the project entries are inert.

Falsifier: a client ≥2.1.211 whose ipe special-cases claudeMdExcludes alongside
fallbackModel. Re-run the grep chain against the installed binary before relying on this.

Fidelity verdict

FIDELITY: GO
Basis: .claude/orchestrator-prompts/arch-v2-context-pipeline-s-e/kickoff.md (rev 5, @ origin/staging)
Round: 5
Audited-SHA: 8fe07fd
Evidence: .claude/orchestrator-prompts/arch-v2-context-pipeline-s-e/kickoff.md — zero PR hunks; git diff --name-only origin/staging 8fe07fd252 -- <kickoff> returns empty (byte-identical to the staging tip), and the whole-PR git diff --name-only origin/staging...8fe07fd252 no longer lists it or docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md — the round-4 MAJOR extra is fully withdrawn. The delta is a single commit (8fe07fd) whose only content is that revert: 2 files, +19/−50, no other surface touched.
Findings: MINOR (extra) — 8 out-of-allowlist files persist at the whole-PR level: tests/install-sh/baselines/{react-native,react-next,react-spa,ts-server}/{brownfield,greenfield}.fingerprint. Not in kickoff §2's permitted set. Carried unchanged from round 4, mechanical regeneration, zero design content — observation only, no rework prescribed. · ADJUDICATED, NOT GRADED (stated so it is not silent) — P2b is absent from the diff while kickoff §1 item 2 still states it as a deliverable. This is the kickoff's OWN prescribed branch, not drift: §1 item 4 binds «verify the client's merge semantics against primary docs in this same task … and PARK if the docs contradict the replace model», and §3a binds «Proceed only on the unambiguous parts». The contradiction is evidenced in-PR at a permitted path — docs/meta-factory/research-patches/2026-08-06-claudemd-overlay-semantics-verdict.md:78-91 (installed binary v2.1.207 commit bc512d563325; ipe unions arrays, fallbackModel the sole replace key) with a version-bump falsifier at :97. Removal residue is recorded at packages/core/hooks/pre-push.ts:35-41 and scripts/worktree-doctor.sh:10-13. Grading this as missing would penalise the diff for obeying the kickoff. · missing: empty · diverged: empty

Round history: r1 STOP → fixed (0eabadc235) · r2 STOP → P2b removed (e94be1e5f9) · r3 REVISE → rework (76d06ce34d) · r4 REVISE → out-of-allowlist hunks dropped (8fe07fd252) · r5 GO (above). The owner-side W-4 spec correction rides PR #1252 — owed on staging independently of this PR (round-5 W-4 note: staging's spec §1.6/§2/§3 still state the replace model until it merges).

Review findings

Watch-list from round 3, pasted unedited per agents/fidelity-auditor.md.

No watch-list from rounds 1-2 was handed to this seat, so ids are assigned fresh this round and no prior-round history is asserted. If an earlier block exists, renumber against it before pasting.

id	criterion	why	defect site	reintroduction tell
W-1	kickoff §3a:216-220 — a fired fork is handed to the operator, never disposed of in-session	the disposition ("remove" vs "re-point at the inverse hazard") is the operator's call; a self-taken one arrives as a fait accompli inside a PR	.claude/orchestrator-prompts/arch-v2-context-pipeline-s-e/kickoff.md:55-56 (this round)	any imperative in an intent artefact that closes a decision the session made about its own scope — "Do not re-open", "settled", "no further discussion", however worded
W-2	kickoff §1 item 2 stays REMOVED unless the union verdict is refuted at the binary	a superset assert under project ∪ local is green by construction — re-adding it ships a gate that can never fail (#hope-as-gate)	packages/core/hooks/pre-push.ts (P2b section, removed this round); scripts/worktree-doctor.sh:10-13	any new pre-push section or doctor arm comparing settings.local.json claudeMdExcludes against the project list in the SUPERSET direction, or a returning AIF_CLAUDEMD_LOCAL_SHADOW_ALLOW
W-3	research-patch §3 falsifier — the union verdict is re-derived from the binary, not from the docs	the docs page produced the WRONG answer twice (§2.3 infers array semantics from an omission); a docs-only re-check re-installs the disproven model	docs/meta-factory/research-patches/2026-08-06-claudemd-overlay-semantics-verdict.md:97-98	an overlay claim citing code.claude.com/docs/en/settings with no accompanying ipe/WSm/Mo grep against an installed claude.exe, or a client-version bump with no re-run recorded
W-4	spec is the design SSOT — a fork correction lands in EVERY place the deliverable is stated	§1.6 says removed while §2/§3 still mandate it; a later stage reading §2 rebuilds P2b and the contradiction is invisible from either half alone	docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md:323-327 and :356 (uncorrected)	a spec whose FORK section carries a dated correction while its §2/§3 prose for the same item is unannotated — grep the item's name and require every hit to be consistent
W-5	kickoff §2:146-159 permitted-files allowlist binds the implementing session	edits to the intent artefact by the session being graded against it are what makes a cold acceptance audit necessary in the first place	.claude/orchestrator-prompts/…-s-e/kickoff.md and docs/superpowers/specs/2026-08-06-…design.md (edited)	any further hunk touching a kickoff, spec, or umbrella charter in a stage PR — including "just fixing the stale line" in the umbrella kickoff flagged as MINOR above
W-6	kickoff §1 item 2 removal is complete at FILE granularity, not section granularity	the section went but its imports stayed; nothing in the toolchain catches an unused import here, so the residue is permanent once merged	packages/core/hooks/pre-push.ts:33 (relative), :34 (execFileSync) — both unused at this SHA	any node:* import in pre-push.ts with zero call sites; re-check with grep -c per named import after any future section removal
W-7	kickoff §3 P3a triple — the gate must be shown RED on the MEMBERSHIP PREDICATE, not on a low ceiling	AIF_ALWAYSON_CEILING=1000 exits 1 with zero work done, so a ceiling-lowering "discrimination" proves nothing about the fix under test	none — preventive (reproduced GREEN this round: 7 KB fixture without paths: → EXIT=1; with paths: → 0)	a discrimination leg quoted as AIF_ALWAYSON_CEILING=<small> rather than a .claude/rules/<fixture>.md pair differing only in ^paths: frontmatter
W-8	kickoff §1 item 3 — host-cc stays UNMEASURED until S-H's host session supplies it	a host ceiling derived from the container is #budget-sized-to-the-wrong-machine and would silently under- or over-gate the operator's real machine	none — preventive (correct at scripts/check-alwayson-budget.sh:35-37)	a second numeric ceiling, an AIF_ALWAYSON_CEILING default change, or a host-cc derivation comment carrying any number at all before S-H lands
W-9	the design-SSOT correction W-4 demands is an OWNER commit, not a stage-PR hunk — W-4 and W-5 are unsatisfiable together, and §2 (the Basis) wins	a watch-list row is continuity, not an amendment to the kickoff; treating W-4 as licence to edit the spec lets a prior seat's finding override the intent artefact's permitted set, and the same hand then writes both the diff and the yardstick	docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md:207-208, 233-242, 323-332, 362 + .claude/orchestrator-prompts/…-s-e/kickoff.md:55-71 (round 4)	a spec/kickoff correction appearing in a stage PR justified by "a previous audit round asked for it"; the conformant shape is the correction in docs/meta-factory/research-patches/* (permitted) plus a separate owner-authored commit for the spec

Round 3: W-1 DEFECT-FOUND · W-2 CLEAN · W-3 CLEAN · W-4 DEFECT-FOUND · W-5 DEFECT-FOUND · W-6 DEFECT-FOUND · W-7 CLEAN · W-8 CLEAN

Round 4: W-1 CLEAN (kickoff.md:55-71 — "Do not re-open" removed; disposition now recorded as an operator-answered fork with both options and consequences at :62-69; the only surviving "settled" is spec:236 "Reading the shipped client settled it", a statement about evidence, not a closure of scope) · W-2 CLEAN (git grep AIF_CLAUDEMD_LOCAL_SHADOW_ALLOW at the audited SHA hits only the historical mention at research-patches/2026-08-06-claudemd-overlay-semantics-verdict.md:123; worktree-doctor.sh:10-13 is a removal comment with no arm; measure-always-on.sh:54-78 implements the UNION direction) · W-3 CLEAN (patch:78 records claude.exe v2.1.207 with the ipe()→WSm()→Mo() extraction, :97 states the ≥2.1.211 falsifier, plus an explicit "Not verified empirically" limit) · W-4 CLEAN (every hit of the withdrawn item is now consistent: spec:207-208 FORK C, :233-242 FORK D, :323-332 struck-through + WITHDRAWN, :362 P2 row "§2 assert 1 … assert 2 WITHDRAWN") · W-5 REINTRODUCED (kickoff.md:55-71 and spec:207-208/233-242/323-332/362 — see the round-4 MAJOR extra) · W-6 CLEAN (pre-push.ts:33 import { resolve, dirname }; execFileSync 0 refs; the two remaining relative hits at :287 and :1485 are prose inside comments; all 8 install-sh fingerprints regenerated) · W-7 N/A — checked as far as cold inputs allow: nothing in this round's diff touches the acceptance-leg surface, and the leg's evidence lives in the PR body, outside the cold seat's inputs; a later seat with PR-body access must verdict it · W-8 CLEAN (check-alwayson-budget.sh:27-38 — single committed ceiling 54,000 B derived from the 48,671 B container baseline × 1.10, host-cc: UNMEASURED — baseline not reachable from the container; S-H's host session supplies it, no host number anywhere) · W-9 REINTRODUCED (row opened this round)

Round 5: W-1 CLEAN (no PR hunk in any intent artefact; the disposition survives only in the permitted research patch, which cites the kickoff's own PARK clause rather than an in-session ruling) · W-2 CLEAN (AIF_CLAUDEMD_LOCAL_SHADOW_ALLOW occurs exactly once in the tree, as prose in the removal record — 2026-08-06-claudemd-overlay-semantics-verdict.md:123; no superset comparison in pre-push.ts or worktree-doctor.sh; scripts/measure-always-on.sh:54-55 implements the union) · W-3 CLEAN (…-verdict.md:78-91 binary-derived _1/ipe/WSm/Mo at v2.1.207 commit bc512d563325; version-bump falsifier at :97) · W-4 N/A (the spec carries no PR hunk this round, by design per W-9; NOTE for the next seat — the owner-side correction is still NOT on origin/staging: newest spec commit is 43bc809, the S-H [W]/[H] verdict, unrelated to overlay semantics, so staging's §1.6 FORK D / §2 item 2 / §3 P2 row still state replace-per-key and still ask for assert 2. The item stays owed OUTSIDE this PR — it rides PR #1252) · W-5 CLEAN (whole-PR file list contains no kickoff, spec, or umbrella charter) · W-6 CLEAN (all 8 imported node:* symbols have call sites in pre-push.ts: existsSync 30, readdirSync 1, readFileSync 3, realpathSync 2, statSync 1, resolve 34, dirname 1, fileURLToPath 2) · W-7 N/A (unchanged reason: the P3a discrimination evidence lives only in the PR body, which the cold-ness constraint bars — unverdictable from these inputs, not unchecked-by-omission; the round-3 in-body reproduction record stands) · W-8 CLEAN (check-alwayson-budget.sh:38 carries ONE committed ceiling, AIF_ALWAYSON_CEILING:-54000, derived at :28-34 from the container baseline 48,671 B × 1.10; the host-cc line at :35-37 carries no number, only "UNMEASURED … S-H's host session supplies it") · W-9 CLEAN (no spec/kickoff hunk justified by a prior audit round; the round-4 prescription was executed by removal, not by re-editing the intent artefacts)

§1.7 Forward-check applied

Checked this round's work against the disciplines that govern it, and let each one change what shipped.

• attention-is-not-a-mechanism.md §1 — the round-3 verdict is recorded as returned (REVISE), not edited toward GO, and the gate stays RED; a body massaged to green would make the acceptance contour a #hope-as-gate. It also supplies the grading language for the removal itself: a superset assert under union is green-by-construction. file:line evidence: .claude/rules/attention-is-not-a-mechanism.md:24
• ai-laziness-traps.md T19 — the audit is a cold seat handed only the staging kickoff and the 3-dot diff, explicitly denied the PR body and comments, and it re-derived the binary claim independently rather than trusting the commit message. file:line evidence: .claude/rules/ai-laziness-traps.md:150
• cold-seat-economy.md §3 — the seat was fresh and narrow with inputs inlined, not a resumed transcript; its watch-list is pasted unedited as the continuity artefact a round-4 seat gets instead of this session's transcript. file:line evidence: .claude/rules/cold-seat-economy.md:1
• phase-research-coverage.md §1.7 — the overlay verdict carries file:line evidence, a stated falsifier (a client ≥2.1.211 whose ipe special-cases claudeMdExcludes), and an explicit note that no live client-level A/B discriminator was obtained. file:line evidence: .claude/rules/phase-research-coverage.md:1
• build-first-reuse-default.md — no new capability: this PR removes a gate, corrects docs and sweeps imports. The one ADOPT verdict in the stage (picomatch, SSOT research(meta-orchestrator-mode-triage): land design doc for /meta-orchestrator extensions #238) is unchanged and still carried by its own commit's trailer. Verdict: compliant.

§1.7 Backward-check applied

Class of this change = a gate removed because its premise was disproven, with the intent artefacts that mandated it edited by the session doing the removing — swept beyond this diff.

• docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md — GAP-FOUND and fixed: §1.6 FORK D said removed while §2:323 and §3:356 still mandated the assert. The general lesson is recorded as W-4: a fork correction must land in every place the deliverable is stated, not only in the FORK section. file:line evidence: docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md:323
• .claude/orchestrator-prompts/arch-v2-context-pipeline/kickoff.md:251 — GAP-FOUND, deliberately NOT patched: the umbrella kickoff still lists the removed deliverable, and it sits outside §2's permitted-files allowlist. Reported rather than fixed, because fixing it repeats the very deviation round 3 graded MAJOR.
• packages/core/hooks/pre-push.ts — swept for other residue of the removed section beyond the named imports; dirname is still used and stays, and grep -n picomatch returns only the guard comment explaining why the package must not come back.
• The prose-vs-identifier sweep gap recurred a third time this stage: an identifier grep found FORK D and missed §2/§3, which state the same requirement in words. Recorded as W-4's reintroduction tell so a later round greps the item's NAME and requires every hit to be consistent.

[artyhoo]

BLOCKED-1 resolved against the shipped client: overlay semantics are MERGE, not REPLACE — P2b is dead machinery

Round 2 flagged that the REPLACE-PER-KEY verdict rested on a mis-attributed docs quote and that the kickoff's PARK was not honoured. That is now settled — not by a better docs citation, but by reading the client that actually runs.

Evidence

Source: the installed binary /opt/homebrew/lib/node_modules/@anthropic-ai/claude-code/bin/claude.exe, v2.1.207, commit bc512d563325 (strings -n 8 + grep). Every symbol in the chain was resolved — nothing inferred:

// fold order over settings sources
_1 = ["userSettings","projectSettings","localSettings","flagSettings","policySettings"]

// source → file mapping
GQ(): projectSettings → .claude/settings.json ; localSettings → .claude/settings.local.json

// Soi() folds each source in through a lodash-mergeWith-shaped call with `ipe` as customizer
Soi(): n = eie(n, sourceSettings, ipe)

function ipe(e,t,r){
  if(Array.isArray(e)&&Array.isArray(t)){
    if(r==="fallbackModel") return t;   // ← the ONLY replace-on-conflict key in the client
    return WSm(e,t);
  }
  return;
}
function WSm(e,t){ return Mo([...e,...t]) }
function Mo(e){ return [...new Set(e)] }

Array settings UNION with dedupe across settings files. fallbackModel is the single hard-coded exception. That is exactly what settings.md:278 documents ("Unlike most array settings, this key does not merge across settings files") — round 2 read the page correctly, and the Info-block quote used in the verdict patch does not support REPLACE.

The binary's own schema text for the key says nothing about non-merging either: "Glob patterns or absolute paths of CLAUDE.md files to exclude from loading. Patterns are matched against absolute file paths using picomatch. Only applies to User, Project, and Local memory types."

Consequence — P2b cannot fire

Under union, the effective list is always project ∪ local ⊇ project. P2b asserts "every file matched by project is also matched by local" — true by construction, so the assert can never go red. A local list can only ADD excludes, never subtract. That makes the section, its escape token, the doctor arm and the meter's local-replace branch dead machinery, exactly as round 2 said.

Decision — PARK honoured, P2b removed

Maintainer decision (2026-08-07): remove, per the kickoff §1 item 4 PARK instruction. Rationale beyond tidiness: a gate that cannot fire is worse than an absent one, because it advertises that the always-on budget is being guarded when it is not.

Surface list — CORRECTED. The first version of this comment listed only the surfaces that
carry a machine-greppable identifier. A prose sweep (git grep -n -I -E 'P2b' plus
REPLACE-PER-KEY|superset|overlay semantics) found three more that name P2b in words. The
corrected list:

Surface	Action
packages/core/hooks/pre-push.ts — section :1287-1430 (localShadowClaudeMdExcludesSection(), header :1287, acceptance note :1309), registry entry :1741-1746, PREPUSH_ONLY dispatch :1814-1815, picomatch module-header note :36-44	remove
scripts/worktree-doctor.sh — the P2b sweep arm, header :82 through :135	remove
scripts/measure-always-on.sh — local-replace branch :60; REPLACE-PER-KEY header comment :13-20	remove branch; rewrite comment to the union model
.claude/orchestrator-prompts/arch-v2-context-pipeline-s-e/kickoff.md — task :55-58, the rev-5 discrimination-pair acceptance :60-67, the replace mandate + PARK instruction :120-124	record that the PARK fired; do not silently drop the task
docs/superpowers/specs/2026-08-06-pipeline-token-economy-design.md §1.6 FORK D — :207, :232-238	correct the FORK D resolution — see below
docs/meta-factory/research-patches/2026-08-06-instructionsloaded-observability-verdict.md:82	drop the stale P2b cross-reference
docs/meta-factory/research-patches/2026-08-06-claudemd-overlay-semantics-verdict.md — :8, :23, :44-48, :83, and the whole §4.2 :85-104	correct in place, do not delete (T18)
packages/core/principles/34-…test.ts (P2a), the picomatch@^4.0.4 pin, SSOT #238	keep — principle 34 imports picomatch directly and is independent of P2b

Verified clean: no CI job guards P2b. The only shadow match under .github/workflows/ is an
unrelated install-sh glob-reach test; the alwayson-budget job is P3a and stays.

The root inversion, for the record

The spec did not merely guess wrong — it overruled a correct draft. …token-economy-design.md:232-234:

"Overlay semantics (round-4 MAJOR-3 — an earlier draft said «project ∪ local», which contradicts P2b): the working model is replace per key … it is the only model under which P2b's superset assert is load-bearing; under union it would be vacuous."

The union model was written down first and correct, then changed because it contradicted the gate
— the design was bent to keep P2b load-bearing rather than the gate re-examined against the client.
The clause "under union it would be vacuous" is the disproof, stated plainly and then treated as a
reason to reject union. Worth carrying past this PR: that shape — "model X would make our gate
pointless, therefore not X" — is the failure, not the wrong docs quote downstream of it.

BLOCKED-2 becomes moot

picomatch appears in pre-push.ts only inside the P2b section (grep -n picomatch packages/core/hooks/pre-push.ts → 36-44 header note, 1298 comment, 1391-1415 all inside the section). Removing P2b removes the package from the shipped hook entirely, so neither the static import nor the 6ec05c0f0d lazy-load workaround is needed. The lazy-load commit stays in history as the record of the pnpm-consumer crash it fixed.

Confidence and falsifier

Falsifier: a client ≥2.1.211 whose ipe special-cases claudeMdExcludes alongside fallbackModel. Re-run the grep chain above against the installed binary before relying on this. The disputed Info block concerns a pre-v2.1.211 migration and the host binary is 2.1.207, i.e. below that line.

Not verified empirically. claude doctor does not print effective settings, so no live A/B discriminator was obtained; the verdict rests on source reading. Note the earlier live "4 local globs work, 3 project relatives do not" observation is non-discriminating — merge and replace predict identical output when the project entries are inert.

Not load-bearing today, but worth stating: the host's .claude/settings.local.json currently has no claudeMdExcludes key at all (jq -c '.claudeMdExcludes' → null), so overlay_source=project and the 48,671 B figure in the PR body is correct as measured. The meter's else-branch is only wrong in the case where a local file does set the key — which is precisely the case P2b was built for.

Side correction to an earlier claim

The normaliser nog() does not drop patterns lacking a leading /. Its if(!n.startsWith("/")) continue only skips the symlink-resolution augmentation; all patterns are returned and matched. Relative entries are inert because picomatch cannot match a relative pattern against an absolute path. Same conclusion as before, different mechanism.

Residual risk this removal does not cover

Under union the real hazard is the inverse one: a local file can silently ADD excludes and hide rules the project expects to be always-on. If that deserves a gate, it needs its own direction and its own justification — not a re-pointing of P2b mid-stage.