chore(corpus): CHAOS-5721 Rescore six refuse rows as any_of under a corpus version bump

[chrisgeo]

Ticket: CHAOS-5721

TEST-EVIDENCE

Scope: data-only change to corpus/corpus.py (7 rows' expect), corpus/test_corpus.py (pin), corpus/README.md (version history). No scorer/engine code touched.

Where the corpus version lives: this corpus has no separate numeric version constant — the corpus version IS this file's git commit sha (git show <sha>:corpus/corpus.py), which is what a scoring run's corpus_version provenance field records (corpus/semantic_verdict.py:build_verdict). This PR's own commit sha is the bump; corpus/README.md's new version-history section says so explicitly rather than implying a constant changed.

BASELINE GUARD disclosure (every row this PR changes; row expectations are chris's data-semantics call):

row id	old expect	new expect	bar movement
basis-discovered-incidents	refuse	any_of{serve(family=discovered_cohort_ranking), refuse}	LOOSENS: a served answer, which always scored disagree under the old scalar refuse, can now score agree/agree_weak via the added serve branch; a refusal still matches (unchanged).
basis-discovered-repo-count	refuse	any_of{serve(family=discovered_cohort_ranking), refuse}	LOOSENS, same shape as above.
cv-org-count-projects	refuse	any_of{serve(family=discovered_cohort_ranking), refuse}	LOOSENS, same shape as above.
basis-grouped-metric-by-repo	refuse	any_of{serve(family=grouped_cohort_status), refuse}	LOOSENS, same shape as above.
basis-grouped-pr-by-project	refuse	any_of{serve(family=grouped_cohort_status), refuse}	LOOSENS, same shape as above.
pos-grouped-per-phrasing	refuse	any_of{serve(family=grouped_cohort_status), refuse}	LOOSENS, same shape as above.
cv-scoped-projects-by-team-bounded	any_of{serve(family=scoped_cohort_status), refuse}	any_of{serve(family=scoped_cohort_status)}	TIGHTENS: a refusal (previously agreeing via the dropped branch, including the observed role-blind declared-kind terminal error) no longer matches any alternative.

Six rows widen acceptance to admit a serve branch that a prior servable-kind gap made unreachable; one row narrows acceptance by removing a refuse branch that had certified a terminal-role error as a match. Old scoring series stays in git history (git show <sha>:corpus/corpus.py), never overwritten; corpus/README.md gets a new ## Version history entry naming both moves. No other row's expect changed.

Red-first (pin fails at origin/main, a9a72b1e): copied origin/main's corpus/corpus.py into this branch's corpus/test_corpus.py and ran it —

$ python3 corpus/test_corpus.py
CorpusValidationError: refuse/decline row set changed since the pinned scalar tranches: added=['basis-discovered-incidents', 'basis-discovered-repo-count', 'basis-grouped-metric-by-repo', 'basis-grouped-pr-by-project', 'cv-org-count-projects', 'pos-grouped-per-phrasing'], removed=[]

(literal message elides the pin dict's own ticket-id substring per this body's one-ticket-token rule; otherwise verbatim.)

Green at this PR's tip (5758d48), fresh pnpm install first (ajv needed by scripts/validate_json_schema.mjs):

$ pnpm install --frozen-lockfile
Done in 1.9s using pnpm v11.21.0

$ python3 corpus/test_corpus.py
PASS: 36 corpus rows validated
PASS: 7 first-tranche + 4 D24 refuse/decline expect values pinned and matched
PASS: 7 this-PR any_of row branch sets pinned and matched
PASS: every corpus row scored except ['cv-named-project-completion', 'neg-single-subject-why']
PASS: baseline sha256 pin matched, 36 baseline rows shape-checked (0 historical-only, not in the current corpus)

PINNED_5721_ANY_OF pins all 7 rows by branch SHAPE (outcome + serve family), never by question text; its own red controls (drifted branch, reverted-to-scalar, deleted row) and green control run inside the same file.

Full pnpm test:corpus (schema validation, semantic_verdict, schema-driven shape generation): rc=0, all PASS lines, no FAIL.

Full ask-dev quality gate, all rc=0:

$ pnpm exec prettier --check .   -> All matched files use Prettier code style!
$ pnpm run lint                  -> eslint . (clean)
$ pnpm run typecheck             -> tsc --noEmit (clean)
$ pnpm run test:unit             -> Test Files 60 passed (60), Tests 977 passed (977)
$ pnpm run build                 -> next build: Compiled successfully, 4/4 pages generated

CI at this PR's tip (5758d48, PR #76): all 12 checks pass — Analyze (actions/javascript-typescript/python), Build, CodeQL, Dependency Audit, E2E smoke, Gitleaks Secret Scan, Quality, Unit tests, run_checks.sh self-test, test.

PR-ref CodeQL at 5758d48, results_count=0 on every analysis:

$ gh api "repos/full-chaos/ask-dev/code-scanning/analyses?ref=refs/pull/76/head"
/language:javascript-typescript  results_count=0
/language:python                 results_count=0
/language:actions                results_count=0

No codex round: data/expectation PR, no scorer code touched.

RISK-NOTES

• Pure data/expectation change: no product, scorer, or engine code touched; expect_schema.py's SCHEMA_VERSION is unchanged (branch shape itself did not change, only which rows use it).
• The rescoring run against these seven rows' new expect is ordered separately, not part of this PR.
• A related in-flight PR changes corpus/semantic_verdict.py only (scoring any_of serve branches from persisted semantic state); this PR does not depend on it and does not touch that file.
• cv-scoped-projects-by-team-bounded's narrowed acceptance means a future rep that still hits the role-blind terminal error will now score as a disagreement instead of silently agreeing — expected and intended by this PR, not a regression to chase.

[linear-code]

CHAOS-5721