compile-clean force-check (2(ii), §1 floor-coverage): DESIGN + MEASURE — registry-leak diagnosis, (A) tree-scoped-builtin design, blast radius 10/102

[gunbai-bot]

Work-item adhoc-6169238c-5e2 (2(ii) compile-clean force-check). Scope confirmed by parent quick-ant-298: DESIGN-FIRST + MEASURE-FIRST, land nothing enforcing. This PR ships only docs/plans/compile-clean-forcecheck.md (no code).

What it establishes (execution-proven, gunbc built from src/v1/stage0)

The compile-clean gate fail-open is NOT "unreached fn bodies escape typecheck" (bodies are always visited). The real mechanism:

1. Registry leak (the brief's utf8_decode_bytes witness). It's referenced at dsl/extdeps/cloud/gcp/secret_manager.dag:71 but defined nowhere in dsl//src/v2; it resolves because it's a hardcoded row in the global builtin_function_registry() (src/v1/04_method.dag:93), an explicitly-marked BRIDGE scaffold. The registry is global, not scoped to the compiled tree → v1-seed intrinsics leak into the substrate compile.
2. Return-type fail-open (separate; ROADMAP §0 line 31 — Enable strict record-field typecheck: close fail-open infer_record_lit hole (§5) #5293 closed only record-field). Split out as (B), its own PR.

(A) direction — tree-scoped builtin availability

Tag each registry row SubstrateAvailable vs SeedOnly; admit SeedOnly only when the entry root is the v1 seed. A substrate compile then fails closed on a seed-only name (the existing "function not found in scope" path). This advances the registry's own marked dissolve-on (04_method.dag:55-62). The enforcing flip is load-bearing (inference scope) + changes what compiles → operator-gated; this doc + the measured number is the sign-off input.

Blast radius (MEASURE-FIRST)

Empty-registry rebuild + real gunbc compile --source-root dsl --source-root src/v2 ... --target rust (throwaway worktree, uncommitted): 102 "function not found in scope" diagnostics, 10 distinct names, all in dsl/, none in src/v2. 8 general primitives (96 sites), filesystem_read (5, known lens-reflection fork), exactly 1 genuine domain leak — utf8_decode_bytes (gcp), already owned by 2(i). So (A)'s rollout is small/tractable, not a flag day. Full table in the doc.

Adjacent: 2(i) (adhoc-8e5771e1-14a) grounds utf8_decode_bytes as a real std fn + deletes its registry row; (A) is what makes that grounding required by construction so the leak can't recur.

🤖 Generated with Claude Code

[gunbai-bot]

CI red investigated — not caused by this PR, and not actionable here.

Failing check: realization_vocab_clean_tree_holds (src/v2/test/claim/realization_vocabulary_containment/clean_tree_test.dag) returned Bool(false) (batch 2 discovery witness, 1 of 637).

This PR's diff is docs-only — a single new file docs/plans/compile-clean-forcecheck.md, no .dag/code/roster changes (git diff --stat origin/main...HEAD = 1 file). A docs-only diff cannot change the evaluation of a .dag witness, so this is an inherited red from main: the #5445 realization_vocabulary_containment hand-frozen roster rotting until warm-lark's roster fix lands (dissolve-on warm-badger-46). Confirmed with the owning lane (quick-ant-298), which explicitly scoped this out: "that's not #5451's content, don't chase it."

No fix pushed (a docs-only PR has nothing to fix here, and a no-op push would only stale the standing approval). Holding frozen at merge-ready pending the inherited red clearing on main + a 2nd review.

— sent from snappy-gull-450

[gunbai-bot]

CI red investigated — inherited from main, not this PR.

Failing witnesses: corpus_live_anchored_modules_clean_holds / corpus_live_clean_tree_holds / extdeps_external_authority_gate_passes (the extdeps external-authority anchor lens, batch 2).

This is a different red from the earlier #5445 one — it appeared after the auto-merge of origin/main (58b2135) pulled in #5452 (std.encoding: declare utf8_decode_bytes ... wire import in secret_manager.dag), which adds dsl/std/encoding.dag and rewires the extdeps module secret_manager.dag that the anchor lens reads live.

Evidence it's inherited: this PR's diff vs main is docs-only — a single file docs/plans/compile-clean-forcecheck.md; the live dsl/extdeps/** tree is byte-identical to main, so this filesystem-read witness returns the same Bool as main. And main is in fact red on it — the #5452 merge commit's own CI run (1bba051f7) concluded failure, as did #5462's.

No fix pushed: a docs-only PR cannot change an extdeps-tree witness, and the fix belongs to #5452's owner (re-green main's anchor lens). Holding frozen; #5451 goes mergeable once main is green again.

— sent from snappy-gull-450

[gunbai-bot]

Correction to my prior comment's attribution (the conclusion is unchanged — still an inherited main red, this docs-only PR can't cause or fix it):

The anchor-lens red is #5429 (warm==cold cache purity adds dsl/extdeps/realization/cache_purity.dag, which doesn't conform to the extdeps anchor gate), not #5452. Bisect (per the owning lane): main was green at 14:44, #5429 merged 15:04 = first failure, #5452 merged 15:17 into already-red main (it only modifies secret_manager.dag with +1 import — can't un-anchor an already-anchored module). So the fix is owned by zesty-crane-172 (conform cache_purity.dag to the anchor gate), not #5452's author. #5451 stays frozen; mergeable once that lands and main re-greens.

— sent from snappy-gull-450