Harden review infrastructure with contextual APIs

[taras]

Why

PR Review and Repo Analysis execute Markdown from the checked-out revision.
#305 is now present on main; this prerequisite keeps its own review documents
compatible with that contract through mechanical props.* spelling changes.
It does not implement or modify the props language behavior.

What changes

• CI runs the checked-out ./dist/xmd inside top-level <Output> regions.
• GitHubAuth.md provides lexically scoped exact-host GitHub authentication.
• ReviewContext, RepositoryInventory, Doctor, and OxlintDiagnostics
  are typed function components with explicit schemas and contextual APIs.
• Doctor classification, Oxlint normalization, comment classification, and
  review state construction use small typed helpers while Markdown remains the
  composition and prompt layer.
• Raw diagnostic payloads do not cross the durable boundary. The review
  boundary fails on malformed or unexpected Oxlint output; published tolerant
  parseDoctorResult and parseDiagnostics behavior remains unchanged.
• Bot replies use GitHub's user.type, and malformed comments payloads fail
  the review instead of becoming empty history.
• Review components whose declared props are consumed in executable Markdown
  use the current props.* binding contract; this is a document migration only.

Diagnostic disposition

The latest review diagnostic report was treated as an engineering signal, not
as a reason to dismiss the review. The new and materially rewritten code now
has decomposed component boundaries, explicit unknown JSON narrowing,
preserved parsing causes, immutable sorting, and corrected nullable/error paths.

The review-only Oxlint configuration excludes conventions incompatible with
this repository's architecture: component filename casing, required named
component exports, export ordering, generator/function style and yield rules,
mechanical import/key ordering and magic-number rules, and readonly parameter
types for contextual component boundaries. The normal repository lint
configuration is unchanged and remains authoritative for those rules. Pure
comment parsing and formatting now lives in a verified code-review-agent
module, leaving CommentReviewData.ts as a thin I/O and orchestration
boundary.

Review guide

Start with: .reviews/ReviewPR.md, .reviews/components/GitHubAuth.md,
.reviews/components/Doctor.ts, and scripts/tests/review-infrastructure.test.ts.

Then review:

1. Doctor.ts, OxlintDiagnostics.ts, ReviewContext.ts, and
   RepositoryInventory.ts for typed contextual boundaries.
2. CommentReviewData.ts, CommentReviewState.ts, and the tolerant parser
   contracts in packages/code-review-agent.
3. .github/workflows/review.yml and repo-analysis.yml for checked-out
   execution and always-uploaded journals.

Look carefully at: exact-host matching, request/body/header preservation,
credential scope, malformed payload failure behavior, and the absence of
journal/result postflight parsing.

What must stay true

• Errors beneath <Output> fail xmd run; ordinary review findings do not.
• No raw Oxlint output, token, arbitrary diagnostic field, or caught error is
  persisted as a document binding.
• GitHub authorization is inherited by projected children, restored after
  scope exit, and absent from durable journal output.
• Function-component generator arguments and published parser behavior remain
  compatible.
• No secret-detection opt-out, bare-prop compatibility fallback, props-language
  implementation change, core change, or runtime change is included.

How to verify it

• deno task lint — pass with 0 errors.
• deno task check — pass, including every .reviews/components/*.ts file.
• pnpm exec tsc --project tsconfig.node.json --noEmit — pass; the ES2022
  Node target no longer depends on Array#toSorted().
• The full Deno suite completed at the preceding rebase head with 410 passed
  and 0 failed; the focused review-infrastructure suite passes 10/10 under
  Deno after the final document migration.
• deno task check:jsr — pass, Success Dry run complete.
• Focused review-infrastructure tests pass under Node and Bun; full local
  portable runs are limited by this environment's loopback socket permission.
  The push-triggered GitHub Node/Bun checks are the authoritative full suites.
• git diff --check — pass.

Scope

Included

• Checked-out review-binary bootstrap and native <Output> failure behavior.
• Scoped GitHub authentication through @effectionx/fetch.
• Typed review components, bounded diagnostics, focused security/failure
  tests, and surgical code-review-agent/Oxlint specification updates.
• Mechanical props.* migrations in the PR-authored review documents required
  by Namespace Markdown props under the props binding #305's now-current language contract.

Intentionally unchanged

• Namespace Markdown props under the props binding #305 props syntax and language implementation; it is already on main.
• WorkflowRun, Workspace restoration, SQLite, replay, Git, Agent, and
  xmd workflow lifecycle code.
• Core and runtime source, generic runtime fetch behavior, public PR model
  types, diff-preview limits, function-component generator arguments, and
  secret-detection defaults.
• Published parser API behavior; no changeset, release metadata, or version
  bump is required.

New dependencies

None.

Risks and limitations

• EnsureOxlint.md retains only irreducible host operations: binary download,
  archive extraction, checksum verification, executable-mode changes, and
  cleanup.
• GitHub CI and the rendered review are being rerun for this correction at the
  exact pushed head.

Scope confirmation

• Every changed file supports review infrastructure.
• Unrelated runtime, lifecycle, and props-language changes are excluded.
• The diagnostic disposition matches the final diff.

[github-actions]

PR #375: Harden review infrastructure with contextual APIs

41 files, +3275 / -2451

Scope

🔴 PR has 5726 lines changed. Split into focused PRs.

🟡 5726 lines changed. PRs under 400 receive more thorough review.

🟡 41 files changed. Are all changes related?

🟡 Changes span 13 directories.

🟡 PR mixes config and source changes.

Structural

✅ No structural bloat detected.

Slop

✅ Slop indicators look low.

Static Analysis

Oxlint: 1 diagnostic across 1 file (1 rule)
Density: 0.000 violations/added-line

no-array-sort (1): packages/code-review-agent/src/doctor.ts

Correctness

No extraneous code patterns detected.