test(headless): split prompt A/B test boundaries

[Astro-Han]

Summary

• Split prompt A/B tests along the production module boundaries instead of keeping generic A/B coverage under the prompt wrapper.
• Move generic summary assertions to ab-summary.test.ts and markdown rendering assertions to ab-render.test.ts.
• Keep prompt-ab-summary.test.ts as a thin prompt id/report-title adapter test.
• Split shared test helpers into focused files: ab-run-fixtures.ts, ab-summary-fixtures.ts, hash-fixture.ts, and temp-dir.ts.

Why

PR #340 intentionally left this as follow-up test organization work. prompt-ab-run.test.ts had grown into a mixed boundary covering generic A/B execution, prompt-specific manifest/resume behavior, source fingerprints, summary/rendering, and prompt run wrapper tests. Splitting it makes future A/B harness changes easier to review without changing production behavior.

Scope

Changed:

• Added focused tests for generic A/B manifest, execution, summary, and markdown rendering.
• Added focused tests for prompt A/B selection, manifest, source fingerprints, summary wrapper, and run wrapper behavior.
• Replaced the old combined A/B fixture helper with smaller run, summary, hash, and temp-dir helpers.
• Updated the hash fixture to generate real sha256:<64 hex> fingerprints.
• Relaxed two resume tests that asserted incidental same-pair A/B start order after pair concurrency; they still assert the same arms are rerun or skipped.
• Added a review-fix commit, 33f1fb02, to address the summary/render and helper-boundary review feedback.

Not included:

• No production code changes.
• No behavior changes to A/B execution, summaries, manifests, or reports.
• No generated benchmark artifacts.

Verification

• node ../../worktree-bootstrap.mjs
• npm --workspace @maka/headless test

Result: 531 headless tests passed.

User-facing impact

None. This is test organization only. No changelog, docs, breaking changes, or migrations are included.

Reviewer notes

The PR now has two commits: the initial test-boundary split, plus a review-fix commit that moves generic summary/render coverage out of the prompt wrapper and splits the helper fixtures by test boundary.