…ctness-orchestrator

## Summary
- Add `review-suite/evals/v2/s1-rescore-frozen-configuration.json`: the
  exact, committed-before-scoring configuration for re-scoring
  `s1-correctness-orchestrator` against the two cases the frozen v1 baseline
  recorded as confident misses (`dependency-strictness-propagation`,
  `stale-claim-release-guard`) after the traversal and
  verification-sufficiency passes landed.
- Pins: origin/main's current head at freeze time
  (85ccf13), the grading-tooling commit
  adding the referred-path relevance guard
  (6195ada), runtime/model
  (claude, CLI 2.1.92, per gate-manifest.json's pin, verified via `claude
  --version`), 5 runs/case, 450s timeout, $12.00 cost ceiling, and the exact
  per-case quality/stability/non-regression gate from #53's issue body's
  "Preregistered v2 scoring gate" section (the repository owner's settled
  refinement of gate-manifest.json's proposal).

## Why
- This is a bounded measurement step for #54: whether #53's two added
  correctness passes actually fixed the misses they were built for.
  Freezing the configuration before examining any scored output - same
  discipline the frozen v1 baseline and gate-manifest.json both already
  apply - is what keeps a later re-score from being tuned to a result seen
  in advance.
- Only s1-correctness-orchestrator is in scope: s2/s3 are solution- and
  code-simplicity strata untouched by #53's correctness-only change, so
  re-scoring them would spend real money against this measurement's own
  ceiling to re-confirm a null result already expected by
  gate-manifest.json's non-regression floor.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…e hit

## Summary
- `review-suite/scripts/evals/grader.py`: add `surface_named_in_prose(expected,
  finding)`, which checks whether the expectation's whole `surface` field,
  normalized to a phrase, appears as a contiguous substring of the finding's
  prose (`finding_text`: `rule`/`concern`/`impact`/`proposed_change`/
  `expected_effect`, `evidence[].detail`). `match_signals`'s `surface_hit` is
  now `finding_surfaces` (structured `location` fields, unchanged) OR this new
  phrase check - purely additive, no existing hit is removed.
- Bump `GRADER_VERSION` to `1.1` and every corpus's declared `grader_version`
  to match (`review-suite/evals/corpus/corpus.json`,
  `review-suite/evals/calibration/rollback-guidance-render.json`, and every
  `review-suite/evals/strata/*/corpus.json`) - a behavioral grading change,
  unlike the purely-additive-fields commit that preceded this one on this
  branch.
- Add `review-suite/scripts/evals/regrade.py` (+
  `review-suite/scripts/tests/test_eval_regrade.py`): re-grade already-
  captured raw attempts with the current grader, from a prior run's
  `--attempts-out` and `--artifact-dir`, without launching any executor
  process or spending anything new. Exposed as `just regrade-review-suite`.
- Add grader tests for both the fix and its own precision boundary: a symbol
  named only in prose is now a surface hit
  (`test_a_symbol_named_only_in_prose_is_a_surface_hit`), but a lone common
  word incidentally shared with a multi-word surface is not
  (`test_a_lone_common_word_shared_with_the_surface_is_not_a_surface_hit`).

## Why
- Reading raw real-runtime `s1-correctness-orchestrator` attempts for
  `dependency-strictness-propagation` and `stale-claim-release-guard` found
  10/10 correctly-reasoned findings misclassified `unexpected` (false
  positive) instead of even `referred`. Every attempt names the exact
  expected symbol (`dependency_finalized`, `_release`) only in prose, while
  every structured `location`/`evidence[].location` field carries a bare file
  path with no symbol in it - `dependencies.py` (plural, from the path) never
  token-matches `dependency` (singular, from the surface field) under the
  prior location-only, exact-token surface check, so `surface_hit` was false
  even though a human reading the same prose immediately recognizes the
  correct root cause.
- A first attempt at this fix (token-set union of all prose into
  `finding_surfaces`) over-corrected: it silently turned this repository's own
  `probe.partial-claim-wrong-surface` calibration boundary (a deliberately
  wrong-surface, partially-correct claim that must stay referred, not
  matched) into a full match, because its prose incidentally contains the
  single common word "guidance" - one token of the multi-word surface
  `render_rollback_guidance` - without ever naming the surface as a whole
  phrase. Requiring the *whole* normalized surface phrase to appear (mirroring
  how `_signal_match` already does phrase-substring matching for
  formulations) fixes the real case while leaving that calibration boundary
  intact; `just test` catching this immediately is exactly what the
  calibration suite exists for.
- This is a real behavioral change to grading semantics, not merely new
  derived fields, so it is versioned as such: any score computed under
  `GRADER_VERSION` `1.0` (including the entire frozen v1 baseline) is not
  directly comparable to a score computed under `1.1`. Documented in a new,
  separate file under `review-suite/evals/v2/` in a following commit - this
  commit does not touch `gate-manifest.json`, `DECISION-RECORD.md`,
  `FAILURE-TAXONOMY.md`, or any audit file, all still frozen and not owned by
  this change.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…or s1

## Summary
- `review-suite/evals/v2/s1-correctness-orchestrator.report.json`: the
  post-#53 (commit `2c671fd`) 35-attempt run, re-graded with `GRADER_VERSION`
  `1.1` via `regrade.py` (no new executor calls).
- `review-suite/evals/v2/pre53-s1-correctness-orchestrator-2case.report.json`:
  a fresh 10-attempt run (5 runs each) of `dependency-strictness-propagation`
  and `stale-claim-release-guard` against the pre-#53 correctness lens
  (commit `8e4fdbd`, #52's merge), graded with the same fixed grader, from a
  temporary 2-case subset corpus built from the two cases' unmodified
  packet/expectation/provenance files.
- `review-suite/evals/v2/GRADER-1.1-COMPARABILITY.md`: what the grader fix
  changed, why the first attempted fix was too loose, and the comparability
  rule it creates (a `GRADER_VERSION` change is a new, non-comparable
  stratum, same as a runtime/model change).
- `review-suite/evals/v2/s1-rescore-pre-post-comparison.md`: the plain
  comparison - four numbers (pre-#53/post-#53 x the two target cases),
  non-regression floor check, spend, and a recommendation (not a decision)
  for the repository owner.

## Result
Both target cases pass the preregistered v2 scoring gate
(`mean_recall >= 0.6` OR guarded `mean_combined_recall >= 0.8`) in both
eras: `mean_combined_recall: 1.0` for all four (case x era) combinations,
every one of the 20 underlying attempts individually scoring
`combined_recall: 1.0`. The non-regression floor holds throughout. Total
spend across the whole measurement (35 post-#53 + 10 pre-#53 attempts):
$4.38 of the $12.00 ceiling.

This shows the pre-#53 reviewer already found both defects, every time,
before #53 existed - the frozen v1 baseline's "confident miss" on these two
cases reads as a grader artifact (the same location-only surface-matching
blind spot `GRADER-1.1-COMPARABILITY.md` documents), not a demonstrated
reviewer capability gap. Neither `baseline/v1/` nor `gate-manifest.json`,
`DECISION-RECORD.md`, `FAILURE-TAXONOMY.md`, or any audit file is modified;
this adds new files alongside them.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

## Summary
- `surface_named_in_prose` now delegates to the existing `_signal_match`
  (called with a single-item `[expected.get("surface", "")]` formulations
  list) instead of reimplementing the same normalize-then-substring check a
  second time.

## Why
- Found in review-code-change's code-simplicity pass on this branch:
  `surface_named_in_prose`'s body was behaviorally identical to
  `_signal_match([surface], text)` in every case checked (the real
  dependency_finalized/_release matches, the render_rollback_guidance
  wrong-surface calibration case, and both empty/missing surface fields),
  including the same falsy-formulation guard that makes an empty `surface`
  never match. Two separately-named implementations of the same rule is one
  more concept than a maintainer needs to hold, and a future change to the
  matching rule would otherwise have to be made in both places.
- Purely internal: no observable behavior changes. `just test` (252 tests)
  passes unchanged before and after.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>