P5 bootstrap-evaluator corpus runtime implementation per worksheet on main (post #4143) — runtime verdict authority, not authoring-time const witnesses

[briansrls]

Summary

Implements the P5 bootstrap-evaluator corpus runtime slice from docs/planning/v4-p5-bootstrap-evaluator-corpus-runtime-worksheet-2026-06-01.md. The manual corpus runner rebuilds each rostered TestClaimEvalSubject with the bootstrap v4_evaluator runtime pin, checks that pin on the execution path, and invokes run_test_claim before folding the existing CorpusEvalReport.

The host transport now fails closed by adding an emitted Rust integration test to the generated crate and running witness_manual_corpus_gate_closed(). The JSON receipt reports runtime_gate_bool: true and runtime_witness: PASS_EXECUTED only after that emitted test passes.

Validation

• bash -n scripts/v4-testclaim-corpus-eval.sh — passed.
• git diff --check — passed.
• CTRL_BUILD_WRAP_CARGO=0 cargo test -p v3-compiler --test integration v4_workflow_ci_runner_dag_smoke_test::v4_workflow_ci_m1_rust_emit_probe_modeled_and_bound_to_ci_yml -- --exact --quiet — passed.
• CTRL_BUILD_WRAP_CARGO=0 cargo test -p v3-compiler --test integration v4_workflow_ci_runner_dag_smoke_test::v4_workflow_ci_bankruptcy_tier0_ -- --quiet — passed.
• CTRL_BUILD_WRAP_CARGO=0 cargo test -p v3-compiler --test integration v4_workflow_ci_runner_dag_smoke_test::v4_workflow_ci_wave1_ -- --quiet — passed.
• CTRL_BUILD_WRAP_CARGO=0 cargo test -p v3-compiler --test integration v4_workflow_ci_runner_dag_smoke_test::v4_workflow_ci_t38_script_checks_generated_manual_corpus_eval_receipt -- --exact --quiet — passed.

Worksheet F1-F6

• F1: script receipt emits execution_status: runtime_verdicts after the emitted Rust runtime gate test passes.
• F2: run_manual_testclaim_corpus_eval executes the modeled runner path via run_test_claim instead of repackaging roster rows.
• F3: runtime subject construction pins bootstrap_projection_inputs.runtime_model.runtime_identity; the runner checks it against v4_evaluator_runtime_id before executing entries.
• F4: manual_corpus_gate(report) reads the runtime-produced CorpusEvalReport; the host transport runs witness_manual_corpus_gate_closed() and fails closed unless it evaluates true.
• F5: the script no longer treats source presence as sufficient; source checks are preflight only, followed by emitted runtime execution.
• F6: dissolution comments are amended in the runner, manual witness, and CI model surfaces.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: abd66b9b · Trigger: schedule
• Thinking: 271s wall

---

BLOCKING (1)

Root Cause

• src/v4/test/claim/workflow/testclaim_corpus_runner.dag BootstrapEvaluatorCorpusRuntimeEval allows any Symbol as evaluator_pin → make the v4 pin invariant unrepresentable or return a report/outcome variant that preserves the mismatch reason and roster facts.

⚠️ The runtime path is mostly clean, but the pin-mismatch branch needs a typed failure shape instead of collapsing to an empty report.

[briansrls]

Violations (could not place on specific lines):

• src/v4/test/claim/workflow/testclaim_corpus_runner.dag:154 BLOCKING: Returning an empty CorpusEvalReport on evaluator-pin mismatch fabricates a plausible empty corpus and drops the roster/pin-mismatch facts instead of carrying a typed failure, violating INVARIANTS.md P2 boundary discipline and facts-flow-forward.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 6c4fdf34 · Trigger: schedule
• Thinking: 238s wall

---

BLOCKING (1)

Root Cause

• src/v4/test/claim/manual/manual_corpus_roster.dag manual_corpus_node_runtime_value_rows remains List<TestClaimRun<Node, RuntimeValue>> → introduce a claim/subject roster carrier that cannot contain authoring-time verdicts, then execute run_test_claim from that carrier.

⚠️ The runtime gate still depends on verdict-bearing roster rows, so the PR has not fully closed the authoring-time const-witness authority split.

[briansrls]

BLOCKING: runtime_subject_from_roster_run still accepts TestClaimRun input, so the runtime path carries and silently drops a pre-authored verdict instead of making the worksheet §4 F2 "claims, not pre-built runs" authority unrepresentable (P2 facts-flow-forward/P5).

[briansrls]

Closing — superseded by #4259 (T-38A.3 manual corpus runner via run_test_claim, MERGED 2026-06-02T01:00:54Z). Original cursor RC addressed at #4259 spawn per operator 2026-06-01 policy. witty-crab-534 was archived; nimble-crab-195 caretaker → neat-wolf adhoc-168ccc84-887 (eager-cat-510) completed the work.

— sent from nimble-dove-733