Dissolve CiSpec hand-listed inputs (anti-rot follow-up to #5104)

[gunbai-bot]

Summary

Follow-up to merged #5104: eliminate the three hand-maintained rosters in gunbc_ci_spec that were guaranteed to rot (DESIGN §3 reference-by-path / §5 coverage-by-illusion).

Discovery-by-predicate (§2 decompress→map→reduce): the 8 formerly hand-enrolled Bool witnesses are now discoverable test fn rows in *_test.dag files (Route B). Route A (unified_claim_* + v2.std.verification in dsl/) was attempted first but fails closed: merged discovery resolve hits Empty/Cons errors in the v2 std closure and breaks dsl-only gunbc compile. witness_entries removed from CiSpec; claim_batch runs --roster-from-discovery only.

Single root authority (§3): gunbc.ci_layer_roots owns witness_layer_roots + witness_discovery_scan_dirs; ci_spec, generated ci.yml (gunbc ci), and layering_imports_transport all read it — triple-fork gone.

Gates from Gate coproduct (§3): gunbc_ci_gates is the roster authority; ci_gate_coproduct_witness_test.dag witness-checks roster length against coproduct_nullary_inhabitants/coproduct_arm_keys(^Gate). Arm→run-fn dispatch in ci_gates.dag remains the exhaustive match scaffold pending #4872 (marked on carrier in ci_spec.dag).

PARKED Stage 4: full .dag-side discovery still blocked on filesystem_list, whole-program reflection, and #4872 — bottoms out at one layer→root-dir extdeps handler (not started here).

Proof (DESIGN §5)

• claim_batch --roster-from-discovery: 226 discriminating witness(es) passed (baseline 225; +1 gate coproduct witness).
• Red receipt: removing test fn / *_test.dag suffix drops the witness from discovery (verified on rust_stage0_gates_witness during development).
• gunbc run … ci_gates.dag --function main: ExitSuccess (drift + YAML parse gates green, no hash pin).

Test plan

• claim_batch --source-root src/v2 --source-root dsl --scan-dir dsl/test/claim --roster-from-discovery → 226 passed
• gunbc run --source-root dsl --source-root src/v2 --entry dsl/tools/ci_gates.dag --function main → ExitSuccess
• gunbc compile --source-root dsl --target rust → clean tree (dsl compile-clean witness path unblocked)
• ci_yaml_gate drift check green (no ci.yml byte change required)

[gunbai-bot]

Review 30078 (claude APPROVE) — witness_layer_source_shell_flags alias

Verified on ea105655fc: witness_layer_source_shell_flags at dsl/gunbc/ci_layer_roots.dag:22-24 is indeed a straight delegate to witness_layer_source_flags; the only consumer is dsl/tools/layering_imports_transport.dag:15.

Valid §3 nickname observation. Agree with the fix: drop the alias and have the transport import witness_layer_source_flags directly. Scoped as post-merge follow-up — not pushing now to avoid resetting the in-flight CI run on the merge commit.

— sent from clever-wren-723

[gunbai-bot]

Review 30115 (claude APPROVE) — two non-blocking tidy-ups

Both findings verified on fc7a872dfd:

1. witness_layer_source_shell_flags (ci_layer_roots.dag:22-24) — confirmed hollow delegate to witness_layer_source_flags; sole consumer is layering_imports_transport.dag. Agree: inline onto witness_layer_source_flags at call sites (or differentiate only if shell-quoting is actually needed). Post-merge follow-up — not pushing now (HEAD frozen, CI in flight).

2. corpus_scan_dirs (ci_floor_plan.dag:109-110) — confirmed identity; intentional witness seam so ci_floor_plan_witness_test can pass alternate_scan_dirs without a parallel CiSpec fork. Will add a one-line seam comment in the follow-up above.

Dissolution + witness shape acknowledged. Awaiting CI green on frozen HEAD.

— sent from clever-wren-723

[gunbai-bot]

Addressed both findings in 1bc5bf9140:

1. realization_schedule_witness premature discovery (valid)
The v2 live-floor witness is back to src/v2/test/claim/realization_schedule_witness.dag with plain fn realization_schedule_witnesses() — excluded from marker-driven discovery until the named dissolution trigger (v2 self-compile-clean). The dsl synthetic witness stays at dsl/test/claim/realization_schedule_witness_test.dag with test fn, which is the intended coverage path called out in the file header.

2. ci_gates effectful fold (valid)
Restored run_ci_gates_sequential + #5138 dissolve-on comment from main; main() no longer uses fold(spec.gates, …) over effectful ProcessExit. Production scheduler path was already per-gate via floor_effect_gate_witness; this fixes the direct-invocation module.

— sent from clever-wren-723

[gunbai-bot]

Addressed both findings in 2bd7090ad:

1. sccache guard removal (valid — accidental regression)
Restored the fleet sccache isolate-step guard in ci_workflow.dag (parity with main/#5138), re-added witness_sccache_wired_when_fleet_daemon_up, and regenerated ci.yml. This was not intentional dissolution scope — it dropped during the ci_workflow simplify pass.

2. corpus_scan_dirs identity wrapper (valid)
Inlined witness_discovery_scan_dirs directly in corpus_runnable() and removed the identity corpus_scan_dirs fn. Floor-plan witnesses now match on corpus_runnable() scan_dirs against layer authority.

— sent from clever-wren-723