Emit the release-build runner through bash_build/bash_emit; delete the dangling compile-gate emitter

[gunbai-bot]

Release-build runner onto bash_build / bash_emit; delete the dangling compile-gate emitter

Covers two (c) rows from #12430's census, ci_release_build_shell_emit_dissolution_trigger and ci_v1_compiler_tests_compile_gate_shell_emit_dissolution_trigger. Node adhoc-da321263-f7f; based on main after #12422.

Release-build (v2.workflow.ci_release_build_emit ci_release_build_script)

• ROOT is gunbc.fleet_release_bins_key root_statement, the same stamp fleet-converge: content-addressed release bins, reused by exact key instead of rebuilt #12422's release-bins steps in this job already use. There is one authority, not a second spelling.
• Artifact verification is gunbc.ci_spec ci_floor_build_verify_stmt, a bash_build command whose paths are $ROOT + /target/release/<name> word concats, folded over ci_floor_required_artifact_names. ci_floor_build_verify_script is now that statement's emission, and artifact_shell_path is deleted.
• sccache report is bash_build_or_else(with_redir(sccache --show-stats, 2>/dev/null), true).
• The EAGAIN retry core is unchanged: it stays the v2.std.orchestration Retry that ci_retry_emit emits, as the brief requires. The three parts join through bash_emit_program_sequence_script.
• A serializer refusal yields a poison marker, which the workflow drift gate reds. Never a partial script.
• Deleted: ci_release_build_shell_emit_scaffold and ci_release_build_shell_emit_dissolution_trigger.

.github/workflows/fleet-converge.yml was regenerated by gunbc run … --entry dag/gunbc/instruments/generated_artifact_gate.dag --function main_wet_one --arg path=.github/workflows/fleet-converge.yml. The diff is 4 lines, all inside the Build release bins step:

line	explanation
run: |- → run: |	The emitted program now ends with a newline (bash_emit terminates every statement), so the YAML block loses its strip indicator. Script bytes are unaffected apart from that trailing newline.
the # 🟡 dissolve-on: ci_release_build_script … comment is removed	Its scaffold and trigger are deleted.
ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd) → ROOT=$('git' 'rev-parse' '--show-toplevel')	Behaviour change, called out as agreed. The || pwd fallback is gone. Outside a git checkout the step now fails, because Actions runs bash -e and the substitution's failure stops the step. It no longer silently uses the working directory. This matches #12422's stamp for the same job. The step always runs after actions/checkout, so there is no difference on the real path.
verify line: words quoted	"$ROOT"'/target/release/claim_executor' '--verify-build-artifacts' "$ROOT"'/target/release/…'. Executed: both lines run against a stub claim_executor that prints its argv, and the 18-word argv is identical (cmp).
sccache --show-stats 2>/dev/null || true → 'sccache' '--show-stats' 2>/dev/null || 'true'	Quoting only.

The build and retry lines are byte-identical. Neither serializer limit is hit: no redirect ends a pipe, and there is no a&&b\|\|c.

Compile gate: deleted, not migrated (parent ruling)

v2.workflow.ci_v1_compiler_tests_compile_gate_emit had no consumer. Its only importer was its own _test, and no workflow carries it: its job was cut on 2026-09-04. Under DESIGN §3c it is dangling, so it is deleted together with its test. Its trigger also cited gunbc.ci_workflow ci_v1_compiler_tests_compile_gate_run_script, which does not exist.

• Rung drop: the loss of that lane is gunbc.rung_drop rust_unit_tests_off_the_merge_path. That row does not name this module, and its restoration trigger is unaffected. A restored lane is authored fresh against whatever route exists then.
• gunbc.ci_spec ci_v1_compiler_tests_compile_gate_note no longer cites the deleted module as the "Operation shape"; it records the deletion.
• Historical mentions are left as they are: the 2026-08-11 correction comment in extdeps.transports.rest, and a past-snapshot path list in gunbc.live_deploy.srv1_residue_rehearsal.

Evidence

• test.claim.gunbc_invoke_witness_test: 7/7 PASS. gunbc_invoke_verify_artifacts_shell_nonempty_holds now pins the exact emitted verify line against an independently built golden.
• build_artifact_verification_witness_test: 3/3 PASS.
• Not addressed here: gunbc.cli_invoke claim_executor_verify_artifacts_shell now has only its witness as a consumer. It has its own dissolve trigger, so it is left for its owner.

🤖 Generated with Claude Code

[briansrls]

APPROVE-MERGE — b73dc31

Reviewed the complete six-file patch, the emitted build-step delta, the shared root/statement vocabulary, the program-sequence serializer, the deleted emitter and its tests, discovered references, and the cited rung-drop authority. No blocking finding.

The replacement is on the production route: release-build imports root_statement, builds verification paths as ROOT-variable plus literal-suffix words from the existing artifact roster, and builds the optional sccache report as a redirected command with an OR fallback. The orchestration-emitted EAGAIN retry core is unchanged. The committed workflow matches that scope: ROOT behavior, quoting, removed scaffold comment and final newline/chomping; build/retry/conditional/timeout are not changed.

Independent narrow execution: I executed reconstructed old/new verify command forms against a temporary recording executable, with ROOT containing spaces and a quote. All 18 words including argv[0] matched byte-for-byte. Under bash -e outside a checkout, the old ROOT form continued through its pwd fallback; the new form stopped with status 128 before the next command. This confirms the declared behavior change, not a regression disguised as formatting. These are shell-form checks, not a local gunbc emission or full build.

The compile-gate emitter's only executable importer found by search is its own deleted test; historical comments/snapshot lists are not runtime consumers. Deleting that dangling pair does not withdraw an executing workflow gate. The cited rust_unit_tests_off_the_merge_path row and its restoration trigger are unchanged. Scope precision: that row explicitly names the v1-compiler --lib population; I am not expanding it into evidence about every separate test crate merely because the removed emitter's name contains 'compiler tests'. No lane-restoration obligation is discharged here.

The verification witness now compares the actual emitted script to independently assembled expected text, retaining the artifact roster and terminating newline, rather than comparing two hand-shell helpers. The reported 10/10 local claims and generated-artifact regeneration remain author-reported; I did not independently execute them. Exact-head PR workflow 36417058603 completed/success.

APPROVE-MERGE at this SHA, through normal exact-head/merge-queue gates. This does not restore or authorize a new CI lane. No merge, enqueue, dispatch or runner mutation performed.