fix(review-fix-loop): unify duplicated JSON-schema validation engine

[shaug]

Summary

Unifies the duplicated generic JSON-schema-subset validation engine
(_path/_is_type/validate_schema) so review-fix-loop's schema validator
reuses the canonical implementation instead of hand-duplicating it:

• Extends the canonical review-suite/scripts/validate.py's _is_type/
  validate_schema to support "type": "integer" and minimum/maximum
  numeric checks, as a backward-compatible superset. No current
  review-packet/review-result schema field uses integer, so this is
  behavior-preserving for every existing consumer (proved by a new test that
  re-runs every existing fixture packet/result through the extended engine and
  asserts identical outcomes).
• Refreshes every bundled references/review-suite/validate.py copy via
  just sync-contracts so the extended engine reaches every consuming skill
  (babysit-pr, implement-ticket, review-code-change,
  review-code-simplicity, review-correctness, review-fix-loop,
  review-solution-simplicity).
• Updates skills/review-fix-loop/scripts/validate.py to import
  _path/_is_type/validate_schema from its bundled
  references/review-suite/validate.py copy via the same
  importlib.util.spec_from_file_location pattern review_gate.py already
  uses for evaluate_bound, instead of hand-duplicating them. Only its own
  schema-specific validate_invocation/validate_checkpoint/
  validate_terminal_result and cross-document checks remain in this file.
• Adds generic-engine unit tests (integer type, minimum, maximum,
  behavior-preservation) to review-suite/scripts/tests/test_contracts.py
  and a regression test in
  skills/review-fix-loop/scripts/tests/test_validate.py proving
  max_fix_cycles non-integer rejection now flows through the shared engine.

This is a follow-up ticket for epic #95, discovered during the epic's own
closeout review: skills/review-fix-loop/scripts/validate.py (added by #96)
contained a near-verbatim copy of the generic engine already canonical in
review-suite/scripts/validate.py. Unlike every other bundled contract copy in
this repository (kept in sync via just sync-contracts and
test_bundled_contracts.py), this duplicated engine had no drift-detection
test.

Non-goals

• No behavior change to review-fix-loop's public validation semantics
  (validate_invocation, validate_checkpoint, validate_terminal_result, or
  the cross-document checks) — this is a pure internal deduplication.
• No change to review-packet/review-result schemas or their existing
  consumers.

Acceptance criteria

• review-suite/scripts/validate.py's generic engine supports "integer"
  type and minimum/maximum checks without changing behavior for any
  existing schema.
• skills/review-fix-loop/scripts/validate.py no longer hand-duplicates
  _path/_is_type/validate_schema; it imports and reuses the canonical
  engine.
• All existing review-fix-loop tests (273) and review-suite tests
  (318) continue to pass unchanged in behavior.
• A regression test demonstrates the shared engine now enforces
  integer/minimum/maximum for at least one review-fix-loop schema
  field (max_fix_cycles).

Validation

• just format — pass, 1 file reformatted (new test wrapped by ruff), no
  behavior change.
• just lint — pass (ruff, mdformat, skills-ref validate on every skill,
  validate_plugins.py).
• just test — full repository suite, pass, 0 failures.
• python3 -m unittest discover -s skills/review-fix-loop/scripts/tests -p 'test_*.py' — 274 tests (273 pre-existing + 1 new), pass.
• python3 -m unittest discover -s review-suite/scripts/tests -p 'test_*.py'
  — 326 tests (318 pre-existing + 8 new), pass.

Review

One full review-code-change pass (fresh, isolated, read-only reviewer given
only raw evidence) returned a clean aggregate verdict on the first pass — no
blocking or strong-recommendation findings from solution-simplicity,
correctness, or code-simplicity. The reviewer independently confirmed: the
canonical engine's additions are structurally inert for every current
review-packet/review-result field (grepped both schemas for integer/
minimum/maximum, found none); the deleted review-fix-loop engine code was
byte-for-byte identical in logic to the canonical engine before this change
(zero behavior drift); all 7 bundled copies are byte-identical to canonical;
and both invocation.schema.json's max_fix_cycles and
terminal-result.schema.json's original_max_fix_cycles declare
{"type": "integer", "minimum": 1, "maximum": 10}, so the new regression test
genuinely exercises the new shared code path.

Guardrail evidence

Diff is 169 insertions / 72 deletions across 12 files — well within
carve-changesets' "a few hundred new or changed lines" preferred range. One
cohesive deliverable (a single internal deduplication), largely mechanical
(bundled-copy resync via just sync-contracts), reviewed clean as one unit on
the first pass. Ordinary single-PR path, no decomposition needed.

Fixes #115
Refs #95