feat: follow-up evidence candidates without automatic recovery (v1.20.0)

[alloevil]

Summary

• Add two exact candidate relationships: same tool differing only in top-level i, and same-turn/same-file explicit modifications.
• Display five result states, field-source evidence, full candidate access and original-result jumps. No automatic failure/recovery or event-membership changes.
• Reject earlier-started parallel calls, missing results/arguments, implicit relative paths and cross-turn file guesses.
• Invalidate affected human review fingerprints when candidate evidence changes; keep storage keys and no-candidate fingerprints compatible.
• Extend synthetic terminal and hosted walkthroughs. Prepare v1.20.0; no dependencies, model calls or new APIs.

Verification

• 255 Node tests / 125 focused tests, 22 new candidate tests pass.
• Normal/demo UI builds, lint with existing findings unchanged, actionlint and 16 automated claims pass.
• Frozen private regression: 15 sessions, 2024 tool results; unchanged 142 failures, 141 pending records, 76 events, 1 recovery. Independently enumerated relationships match; 3 only-i candidates are failure/running/success. No raw logs or detailed private artifacts included.
• Browser: all five states, five-item pagination, unrelated-file/turn exclusion, source-result jump, append-triggered stale review, narrow layout and static-demo zero API requests pass.
• Staged secret scan passes; private output remains ignored.

Boundaries

Candidate success is not task verification; same-file relation requires explicit path context and does not prove causality. Old notes on affected events require review. Not a time-saving or global accuracy claim. Publish only after protected checks.