fix(review): bound manifest path instruction globs

[JSONbored]

Motivation

• The review.path_instructions[].path field was accepted without per-item length/character limits and was repeatedly compiled into regexes and emitted verbatim into the AI prompt, enabling CPU/memory/AI-budget exhaustion from malicious manifests.
• The change hardens manifest parsing and AI-review prompt construction so repo-controlled globs cannot trigger repeated expensive regex compilation or unbounded model input.

Description

• Added MAX_REVIEW_PATH_GUIDANCE_LENGTH and CONTROL_CHARACTER_PATTERN and validate review.path_instructions[].path to reject control characters and truncate to MAX_ITEM_LENGTH during parsing in parseReviewPathInstructions in src/signals/focus-manifest.ts.
• Changed resolveReviewPathInstructions to normalize changed paths once, compile each manifest glob once via compileManifestPathMatcher, and evaluate those matchers against normalized paths rather than recompiling per path.
• Bounded the assembled AI-review guidance section to a safe character budget (4,000 chars) and include an omission note when entries are skipped to keep the final prompt size constrained.
• Added unit tests in test/unit/focus-manifest.test.ts to cover control-character rejection, overlong-path truncation, prompt guidance bounding, and blank-normalized-path behavior.

Testing

• Ran npx vitest run test/unit/focus-manifest.test.ts and all tests in that file passed (113 tests) covering the new branches and regressions.
• Ran npm run typecheck and tsc --noEmit which succeeded with no type errors.
• Attempted coverage: npx vitest --coverage for the unit file and npm run test:coverage for the full suite executed tests, but coverage generation failed due to an unrelated coverage provider/runtime error (TypeError: jsTokens is not a function) coming from the coverage toolchain; this prevented producing a full coverage report locally.
• npm audit --audit-level=moderate could not complete due to the registry returning 403 Forbidden in this environment.

---

Codex Task

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.