fix(review): remember resolved review threads

[Sayt-0]

Summary

• persist marked bot review-thread history before diff-related early exits
• suppress equivalent findings from current resolved or unresolved threads while reassessing outdated threads
• prevent silent severity escalation and calibrate external factual claims in the review policy
• extend deterministic deduplication across REST comments and GraphQL thread state
• document the reviewer memory model and add regressions based on docker/sandboxes#4890

Closes #81.

Issue expectations

Expectation	Implementation
Resolved finding on unchanged code does not resurface	Current resolved marked bot threads suppress deterministic duplicates
Current unresolved duplicates are not reposted	Current unresolved marked bot threads use the same suppression path
Changed or outdated code can be reassessed	Explicitly outdated GraphQL roots are excluded from legacy REST dedupe by correlated comment ID
No silent severity escalation	Unchanged current duplicates are suppressed; changed-code escalation requires explicit new evidence and is audited
External claims are calibrated	Platform, version, and library claims require authoritative evidence or conditional wording
Full immutable SHA semantics are accurate	Policy distinguishes content integrity from availability under history rewrites
Memory behavior is visible	README, posting-format guidance, and contributor documentation describe the lifecycle rules

Validation

• pnpm test: 934 passed across 33 files
• pnpm typecheck
• pnpm exec biome ci .
• pnpm build
• actionlint -shellcheck=
• tests/test-job-summary.sh
• tests/test-output-extraction.sh
• tests/test-release-caller-permissions.sh
• main integration suite: 15 passed
• git diff --check

The GitHub credentials integration suite was not completed locally because it waits on external 1Password credentials. The PR-agent model evaluation was not run because no model API credential was available in the local environment.