[long-run] Run the corrected long-run promotion gate

[chaoz23]

Parent: #215. Depends on #216, #217, and #218. Also depends on chaoz23/inkbench#3, chaoz23/inkbench#4, and chaoz23/inkbench#5; freeze chaoz23/inkbench#6 and chaoz23/inkbench#7 before execution.

Product outcome

Determine whether the candidate converts a matched creator-owned resource grant into more sustained actionable value—or stops materially earlier and more honestly—without sacrificing InkCheck's early discovery or witness integrity.

Preregistered protocol

• Freeze exact InkBench/adapters/analysis, dependency lock, Node/V8, compiler, baseline/candidate CLI bytes, fixtures/manifests, machine posture, serial order, endpoints, exclusions, and stopping rules.
• Reject mixed/legacy fingerprints and do not resume the old 60-minute matrix.
• Primary arms: pinned b2651f5 one-core portfolio, candidate with only the long-run change, and simple systematic/coverage controls. Keep product auto-concurrency as a labeled sensitivity arm.
• Core families: rare-prefix, novelty-honeypot, compound-needle, false-novelty, delayed-consequence, order-dependent, revisit-after-mutation, and deep-corridor.
• Add multi-bug Intercept-20/Heresy-II-30 and clean public The Intercept, Heresy II, and Dog Ink Adventure cases.
• Use at least six held-out structural variants × five paired search seeds per generated family; fixtures are independent clusters and seeds are nested.
• Run fresh 20-minute cells first; run a fresh 60-minute matrix only after identity, witness, stop, adapter, and telemetry audits pass.

Acceptance criteria

• Zero lost/replay-invalid critical evidence, hard OOMs, malformed partial reports, mixed fingerprints, silent limit violations, or duplicate logical cells.
• No planted bug loss; one-sided 95% upper bound for paired time-to-first-discovery ratio is no worse than 1.10 per required family.
• Candidate satisfies at least one preregistered material-value path: new critical evidence on two independent fixtures; at least 25% lower retained GiB-minutes/cost at equal critical yield; at least 2× longer useful operation plus 25% better category-specific efficiency; or at least 25% cost saved by an earlier honest partial stop with no common-horizon evidence loss.
• Terminal-state multiplicity alone cannot pass.
• Compact shared checkpoints before frontier artifacts dominate campaigns #156's 500K Intercept-under-512-MiB or matched-value gate and exact split equivalence remain true.
• Adapt first result windows to observed throughput and resource growth #190's useful reopenable result before 50% of Dog/Heresy time or memory envelope remains true.
• Launched-cell, fixed-grant-complete, and observed-anytime outcomes are separate; resource stops use competing-risk/common-horizon analysis rather than naive independent censoring.
• Full machine-readable artifacts, negative results, and family-level paired estimates are published.

Non-goals

• General algorithm-superiority claims.
• Retrofitting the stopped legacy matrix.
• Promoting a policy on terminal diversity alone.

[chaoz23]

Proof-gate checkpoint — 2026-08-25

No new InkBench matrix has started. Public InkCheck main is 52e1721; the historical evidence source/executable remain preserved, and PR #227's 50K calibration is only a narrow overhead/parity check.

Before execution: finish #156/#216/#190/#217/#218, freeze the candidate source/tree/CLI bytes/policy, add the InkBench bridge that preserves privacy-safe owner/epoch receipts and effective settings, finish the operative InkBench #3/#4 validity gaps, and freeze #6/#7 topology/instrumentation decisions. The first eligible proof is a fresh audited serial 20-minute matrix; the 60-minute tier follows only after the mechanical and methodology audit passes. Never resume the stopped historical matrix. This issue remains open.