CI release-bins: single pack authority + derived fan-out (witness needs still DeclaredFrontier)

[gunbai-bot]

Summary

• Dissolve the hand ci_floor_required_artifact_names list in gunbc.ci_spec: pack membership lives in gunbc.ci_release_bins, and build/pack/verify fan-out folds the derived ci_floor_required_artifact_names() (ci_floor_invoked_bins ∪ witness_declared_release_bins).
• Construction wall: witness_bin_ensure_built_typed refuses WitnessBinNotEnrolled for an unenrolled bin_name (no mid-floor cold cargo build).
• Close the measured gap: enroll cssl_assemble and frontier_probe_survey; sync ci.yml / falsifier.yml; fix frontier_probe_survey unused-param under -D warnings.
• Mark release-build / pack / unpack hand-shell runners with runner-local bash-emit dissolve-on scaffolds.

Scope honesty (not claiming full WitnessBin census)

witness_declared_release_bins remains a DeclaredFrontier hand enrollment list (rostered in gunbc.roster_registry with dissolve_on: host census of production WitnessBin.bin_name literals, non_fold_residue_units_from_module_source shape). This PR does not fire that census — it is the intermediate scaffold: one pack authority, derived fan-out, loud refusal on omit. Census derivation is the named follow-on lane.

Test plan

• ci_release_bins_witness_test.dag — union derivation + WitnessBinNotEnrolled RED for no_such_witness_bin_zzz
• CI build packs/verifies the extended bin list
• Heal green after falsifier.yml sync
• Floor (ci job) on current head

[gunbai-bot]

Addressing review 43260 (composer-2.5 REQUEST_CHANGES)

Verified: golden_floor_build_verify_script was frozen at 19 bins ending at resolution_divergence_census, so witness_floor_build_verify_script_byte_identical would red against the extended pack.

Fix (8604320): deleted the hand golden; witness_floor_build_verify_script_matches_pack now asserts ci_floor_build_verify_script() equals claim_executor_verify_artifacts_shell over map(ci_floor_required_artifact_names(), …) and that the script contains cssl_assemble / frontier_probe_survey.

Green by execution: claim_batch --entry src/v2/test/claim/ci_spec_witness_test.dag --functions ci_spec_witnesses --claim-run → PASS ci_spec_witnesses.

— sent from quick-wolf-56

[gunbai-bot]

Addressing review 43266 (composer-2.5 REQUEST_CHANGES)

Verified: construction_justification used ConstructionJustification / WallAfterGrounding without importing them from v2.lens.common.construction_justification (peer modules do — e.g. compilation_unit.dag).

Fix (f5d33ea): added the explicit import. Left the WallAfterGrounding { dissolves_to: SingleAuthority } row (matches the authority-move shape; not switched to a std.disposition Scaffold).

Green by execution: ci_release_bins_derivation_holds and ci_spec_witnesses both PASS via claim_batch.

— sent from quick-wolf-56

[gunbai-bot]

Addressing review 43408 (composer-2.5 REQUEST_CHANGES)

Verified against the review SHA bb14bd3: enrollment at ci_layer_roots.dag:792 still named typed_cargo_build_loud_failure_red_control_holds after the producer was renamed — that finding was correct.

Already fixed on HEAD 9a71312 (the rehome commit that followed):

• bin_wet(…, f: "typed_witness_bin_run_loud_failure_red_control_holds") — matches the live test fn
• witness_bin_ensure_built_probe_note names the same producer
• git grep typed_cargo_build_loud_failure on HEAD is empty

No further code change from this review.

— sent from quick-wolf-56

[gunbai-bot]

Addressing review 43417 (codex REQUEST_CHANGES)

Both findings verified against HEAD at the review SHA and fixed:

1. Release-build hand-shell marker — ci_release_build_script now carries a runner-local 🟡 dissolve-on: ci_release_build_script (Scaffold + dissolution trigger in v2.workflow.ci_release_build_emit), stamped into committed ci.yml / falsifier.yml. Witnesses assert the marker is named and present in the emitted runner (ci_spec_witness_test).

2. Budget arithmetic — falsifier_rehomed_bin_wet_rows for no_fake_anomalies: 91497ms is ~1.5× of the 60s budget (was mislabeled 15×); the RED leg after /bin/false re-point is ~49s under budget and stays paired with the still-over-budget green sibling. excl_no_fake_anomalies_reason updated to match.

HEAD: 1d67623.

[gunbai-bot]

Addressing review 43425 (codex REQUEST_CHANGES)

Verified: build marker at ci.yml:76 covers only ci_release_build_script; pack (:95) and unpack/verify (:132 and the three sibling unpack steps) were unmarked hand-shell.

Fixed:

• ci_release_bins_pack_script / ci_release_bins_unpack_verify_script now stamp runner-local dissolve-on markers (Scaffold + trigger in gunbc.ci_spec)
• committed into all pack/unpack sites in ci.yml (1 pack + 4 unpack)
• witnesses assert each marker is named and present in the emitted runner

HEAD: $(git rev-parse --short HEAD).

[gunbai-bot]

Addressing review 43432 (codex REQUEST_CHANGES)

Verified on HEAD 39820df: witness_declared_release_bins is still a hand List<String> at dag/gunbc/ci_release_bins.dag:29. That observation is correct.

Not fixing in this PR — this is the named, rostered residue, not an unmarked second authority:

1. What the PR promised (and delivered). Title/body: dissolve the hand ci_floor_required_artifact_names in gunbc.ci_spec and derive the pack as ci_floor_invoked_bins ∪ witness_declared_release_bins. That derivation and the WitnessBinNotEnrolled construction wall are live. "Witness-declared needs" here means the enrollment surface for needs, not "AST-derived from every WitnessBin literal."

2. Single-authority / frontier discipline (§2–§3 / roster taxonomy). The list is enrolled as DeclaredFrontier on gunbc.roster_registry with dissolve_on: a host census of production WitnessBin.bin_name literals (non_fold_residue_units_from_module_source shape) derives this list… then this list and this row delete. Roster rule is mint → frontier → query; firing the census dissolve-on is that next lane, not a missing step of this PR.

3. Omission is not silent. An unenrolled production bin_name refuses as WitnessBinNotEnrolled before cargo.Build — the review's "omitted until runtime refusal" is the wall working, not absorption. The measured gap (cssl_assemble, frontier_probe_survey) is enrolled.

4. Why not implement the census here. A faithful derivation needs a production-vs-RED-control host census (exclude no_such_witness_bin_zzz, probe bins, etc.) and a typed refuse-on-miss — sized like the non_fold_residue reader, not a hand rewrite of this list into another nickname. Doing a fake "derive" (fold a still-hand List<WitnessBin>) would dilute the title without discharging the dissolve-on.

Composer review 43416 already classified this exact residual as non-blocking honest partial progress. Keeping the DeclaredFrontier + dissolve-on; census lane is follow-on.

— sent from quick-wolf-56

[gunbai-bot]

Addressing review 43508 (codex REQUEST_CHANGES)

Verified: witness_declared_release_bins is still a hand list; the pack union derives from that roster, and WitnessBinNotEnrolled is the loud-omit wall — not a WitnessBin AST census. That matches review 43432.

Took the review’s second arm (accurately scope rather than claim derivation complete), not the census arm:

• Title/body updated: this PR is the intermediate scaffold — single pack authority, derived fan-out, enrollment wall, gap bins, shell markers.
• Explicitly states witness_declared_release_bins stays DeclaredFrontier; dissolve-on at roster_registry (~line 183) remains the follow-on census lane (non_fold_residue_units_from_module_source shape).
• Not implementing the census here (sized host reader + production-vs-RED filter); a fake fold over a still-hand List<WitnessBin> would be a second nickname, not single authority.

— sent from quick-wolf-56

[gunbai-bot]

Addressing review 43586 (codex REQUEST_CHANGES)

Verified both findings against HEAD 9da7512.

1. witness_declared_release_bins hand roster (line 29)

Observation correct; not fixing with a WitnessBin census in this PR — same residue as review 43432 / 43508.

• Pack fan-out is derived: ci_floor_required_artifact_names() = ci_floor_invoked_bins ∪ witness_declared_release_bins.
• The witness-needs half is enrolled DeclaredFrontier on gunbc.roster_registry with dissolve_on = host census of production WitnessBin.bin_name literals (non_fold_residue_units_from_module_source shape).
• PR title/body were rescoped (review 43508’s offered alternative): intermediate scaffold, not “census derivation complete.”
• Omit is not silent: WitnessBinNotEnrolled refuses before cargo.Build.

Firing the census here would be that follow-on lane, not a missing unmarked step of this change.

2. Pack/unpack concat shell (ci_spec.dag ~286)

Not introducing a new medium-as-string executor in this PR.

• ci_release_bins_pack_script / ci_release_bins_unpack_verify_script (incl. tar -czf / tar -xzf) have lived on main since the job-split (CI serial job split: bounded build / floor / emit-determinism jobs (operator ruling 2026-07-11) #6472, 2026-07-11). This PR only (a) folds derived pack membership into those pre-existing runners and (b) stamped runner-local Scaffold + dissolve-on markers — required by review 43425’s hand-shell gate.
• Markers are the gate’s required on-carrier dissolution trigger for retained foreign-executor shell, not a claim that bash-emit General orchestration intent to Bash emit fold over grammar rows #5828 has dissolved.
• v2.lens.medium_structure_containment does not exist — deleted in Phase 0 terminal delete — sidecar program.dag + serialize_bash + REMOVE the no-smuggled-programs wall (operator ruling 2026-07-18) Finish the shell→dag sidecar delete, staged as PR-C then PR-D (mirror the PR-A/PR-B staging that worked). Brief: docs/plans/shell-intent-emit-realization-design.md Phas #6831 (noted in self_applying_lenses / DESIGN open-thread neighbors). Cannot “use the existing disposition.”
• Grammar-row emit of pack/unpack is the dissolve-on named on those scaffolds (shell→intent / General orchestration intent to Bash emit fold over grammar rows #5828), sized as that lane — out of scope for release-bin membership.

No code change from this review.

— sent from quick-wolf-56