research(ci): prove a fast-first tier without weakening the final required suite

[BaseInfinity]

Feedback type

Process improvement from a consumer repository.

Consumer evidence

BaseInfinity/m180-jumpseat PR #111 changed planning documentation and lifecycle contract tests only. Its focused document-learning/roadmap gate passed 36/36 in about five seconds, while the unchanged complete hosted required suite passed 2,141/2,141 in 22m15s. The full gate added valuable clean-Windows confidence, but a cheap contract failure should be visible before spending the entire runner duration.

The same consumer needed a lifecycle-only PR #104 after terminal reconciliation; that seven-file documentation/test change also ran the complete 2,136-test suite locally and in hosted CI.

No source code, private evidence, secrets, or business logic was scanned for this report.

Desired behavior

Research and prove an opt-in fast-first, full-final CI contract:

1. Run deterministic changed-surface contracts first (for example docs/lifecycle structure, generated-state verification, or directly affected unit tests).
2. Fail quickly and retain a concise diagnostic when that tier is red.
3. Keep the repository's complete required suite as the final merge gate; the fast tier must not certify the PR by itself.
4. Treat cross-cutting files, proof machinery, hooks, dependency manifests, and ambiguous changes as full-suite work immediately.
5. Record which tier ran, why it was selected, elapsed time, and the exact final gate result.

Acceptance criteria

• Build synthetic fixtures for docs-only, lifecycle-only, code, hook/proof, dependency, and ambiguous diffs.
• Prove docs/lifecycle defects fail in the fast tier before the complete suite starts.
• Prove every merge candidate still receives the declared complete required gate.
• Prove changed-file classification cannot skip the final gate or silently downgrade ambiguous/cross-cutting changes.
• Compare wall-clock and runner-time savings against the existing single broad job.
• Preserve customized consumer workflows; adoption is opt-in.
• Emit retained failure summaries instead of flooding or truncating the useful diagnostic.

Related

• Audit CI/CD for independent value, overlap, and release enforcement #108 — CI/CD independent-value and overlap audit
• bug: retain failed-test summaries when broad proof output is truncated #124 — retained failed-test summaries
• BaseInfinity/m180-jumpseat PR Audit and bind final review approval to the exact integrated candidate #111 — 36/36 focused versus 2,141/2,141 hosted in 22m15s
• BaseInfinity/claude-sdlc-harness corresponding consumer-pattern issue

[BaseInfinity]

Fresh #107 timing evidence: avoid a second full run for lifecycle-only reconciliation

BaseInfinity/m180-jumpseat supplied a useful paired comparison from one issue boundary:

• Candidate PR Reviewers reconcile with EACH OTHER, verbatim: independent → cross-feed → argue → one answer #116: mutation-authority/proof machinery; hosted 2,209-test gate passed in 17m38s. This is the right place for the complete clean-Windows gate.
• Lifecycle PR fix: bound SDLC review and repair loops #117: six documentation/lifecycle-contract files, 26 insertions / 22 deletions. Focused affected checks passed 44/44 in 2.83s; the complete local suite passed 2,209/2,209 in 248.84s; the unchanged complete hosted gate passed again in 22m58s.

The second hosted run added little distinct evidence after #116's frozen issue candidate was already green and merged. It existed only to change candidate wording/counts to terminal PR/issue/milestone state.

Please measure two policies, not only job ordering:

1. Fast-first/full-final: affected contracts first, complete clean run on the frozen issue/integration candidate.
2. Lifecycle terminal follow-up: fail-closed changed-path plus lifecycle/dashboard contracts, with no second full 2,209-test run when the implementation tree is already the exact green merged ancestor and only declarative terminal state changes.

Unknown paths, code, hooks/proof, dependencies, security/auth, mutation authority, packaging, releases, or a non-ancestor implementation tree must still route to the full gate. Record the ancestor/tree reason in machine-readable evidence.

This is additive evidence for #139/#108, not a request to remove clean-machine proof.