Implement review-fix-loop reviewer isolation and complete-review orchestration

[shaug]

Outcome

Implement complete-review orchestration with fresh read-only reviewer subagents by default and an explicit in-agent override.

Parent epic: #95
Design: design/review-fix-loop.md

Scope

• Resolve requested review lenses and require a complete review result for each cycle.
• Use a fresh aggregate-review subagent by default when the host supports subagents.
• Allow nested lenses to share that aggregate-review subagent.
• Reinforce read-only behavior in prompts, allowed operations, evidence collection, and result validation.
• Support in-agent review only through an explicit invocation override.
• Normalize findings for deterministic selection and checkpointing.

Non-goals

• Building new review lenses that are not already available.
• Applying fixes inside reviewer agents.
• Locking, promotion, or publication.

Acceptance criteria

• Default-capable hosts create a fresh reviewer subagent for every complete review.
• Reviewer instructions explicitly prohibit worktree mutation and implementation.
• Reviewer output is rejected if required lenses or evidence are incomplete.
• Nested lens execution can share one aggregate-review subagent without weakening completeness.
• In-agent execution occurs only when explicitly requested and is recorded in the result.
• Tests detect attempted reviewer mutation and fail the cycle closed.

Validation

• Unit and integration tests for lens resolution, reviewer freshness, mutation detection, incomplete output, and explicit override behavior.
• Repository just format, just lint, and just test.

Dependencies

Blocked by the contract leaf.

[shaug]

Implemented, reviewed, and merged as PR #110 (squash commit 8ef6db199526beb6e44408f3848f679dd88edb58 on main, direct child of the merged #97 base 26b4cf47168dc8432f7d6e5e4597439af6391a51).

Acceptance criteria

• Default-capable hosts create a fresh reviewer subagent for every complete review — resolve_review_execution_mode defaults to fresh_subagent with no automatic fallback (blocked/missing_capability when unsupported).
• Reviewer instructions explicitly prohibit worktree mutation and implementation — REVIEWER_PROHIBITIONS / build_reviewer_briefing, tested.
• Reviewer output is rejected if required lenses or evidence are incomplete — evaluate_review_result (lens completeness for clean via the bundled contract's own lens-execution check; packet-evidence completeness via _check_clean_requires_passing_validation).
• Nested lens execution can share one aggregate-review subagent without weakening completeness — one review_records entry per aggregate pass regardless of nested lens count; completeness still enforced by the bundled contract.
• In-agent execution occurs only when explicitly requested and is recorded in the result — resolve_review_execution_mode, review_independence field.
• Tests detect attempted reviewer mutation and fail the cycle closed — detect_worktree_mutation (including refs, excluding refs/remotes/*) forces write_isolation: "violated", and validate_terminal_result rejects converged for any review record with a mutation attempt (end-to-end test).

Validation

• python3 -m unittest discover -s skills/review-fix-loop/scripts/tests -p 'test_*.py' — 232 tests pass
• just test — all skill suites and review-suite/scripts/tests (318 tests) pass
• just lint — ruff, mdformat, skills-ref validate (all 9 skills), plugin packaging validation pass
• just format — no diffs
• Repository-owned review-code-change: clean aggregate (all three lenses fresh and clean) at the exact merged head, after 7 fix/re-review cycles (refs-mutation detection, a contract-legal blocked-result rejection, duplicated binding logic, an ignored-worktree false positive, fail-open snapshot validation, a dead packet-less evaluator path, a rebase-introduced changelog SHA issue, and prose/test verbosity — all found and fixed)
• Post-merge: worktree tree is byte-identical to the merged main tip; full test suite re-run green on that exact tree

Non-goals confirmed out of scope (per design)

New review lenses, applying fixes inside reviewer agents, and locking/promotion/worktree management/publication (issue #97, merged separately) were correctly not built here.

Closing per the epic #95 completion policy (merge after gates plus manual tracker-transition authority).