feat: automatic execution health and modification/check chronology (v1.21.0)

[alloevil]

Direction

Automatic execution evidence first, not a required manual review queue. Preserve existing diagnostics and review data.

Changes

• Default factual health with call-result gaps and unique candidate counts; manual notes/transfers are opt-in and never hide automatic events.
• Compare explicit modification-tool successes to earlier, overlapping and later check-call records; all evidence remains navigable.
• Bounded command lexer recognizes test runner/script-name segments; compound/pipeline fragment outcome stays unknown regardless of wrapper success.
• Known-directory exclusions, simultaneous-call ambiguity, missing/orphan IDs and unsupported coverage are explicit.
• Synthetic temporal demo and revised roadmap/docs; v1.21.0 version and claims aligned.

Verification

• 290 Node tests pass, including 200 fixed health transformations and 120 temporal transformations; no human labels/model judge.
• Normal/demo builds, lint (existing findings unchanged), actionlint and 16 automatic claims pass.
• Frozen 15-session/2024-result regression: exact v1.20 diagnostic output unchanged. 59 check-containing calls identified, all fragment outcomes unknown; raw logs remain private.
• Browser: default zero review-key reads, storage denial isolation, optional-note preservation, live result gaps, before/overlap/after sequence and desktop/mobile source jumps; static demo makes zero API requests.
• Staged secret scan and package exclusion checks pass.

Limits

No file-coverage/task-success proof, no implicit shell/cwd interpretation, no atomic review changes or new dependencies. Publish only after protected checks.