fix: improve completion marker compliance

[willytop8]

Summary

Improve textual completion-marker compliance without loosening the strict detector or adjacent-evidence gate.

Changes

• show [goal:evidence] and [goal:complete] as consecutive plain-text lines in the continuation prompt
• explicitly prohibit Markdown/backtick wrapping and blank-line separation
• keep normal and budget-wrapup continuation prompts within their existing size limits
• add focused regression assertions and an Unreleased changelog entry

Evidence

A real-provider prompt matrix reproduced common failures with the previous wording: only 2/10 models produced evidence and completion accepted by the production parser. A longer candidate improved this to 8/10. The final compact worktree wording produced canonical adjacent protocol lines in all seven targeted baseline-failure models.

The marker detector and evidence extractor are unchanged.

Testing

• npm test — 249/249 pass
• npm run release:check — pass
• coverage — 98.47% lines / 90.50% branches / 89.98% functions
• type contract — NodeNext + Bundler pass
• mutation contract — 5/5 critical mutants killed
• behavior benchmark — 100/100
• packed host/tool contracts — pass, exact 11-tool surface
• installer verifier — 6/6, all 7 hooks
• production dependency audit — 0 vulnerabilities
• provider validation — canonical adjacent completion protocol in 7/7 targeted models

Checklist

• Tests added/updated
• Changelog updated
• No parser relaxation
• No breaking API changes
• Full local release gate passes