Closes #282

What

mutmut categorically skips the body of any @dataclass-decorated
ClassDef (mutmut/mutation/file_mutation.py:236). Issues #262/#280/#283
established the fix — extract methods to free functions, keep the
dataclass pure data — for ConstraintSet, RepoBadge, and Ranking.
This PR finishes the systemic sweep issue #282 found: all 30 remaining
@dataclass-with-methods classes across mcp_server/, fuzz/, and
benchmarks/lib/ (26 source files), each method extracted to a free
function taking the dataclass as its first argument, call sites updated,
existing behavior preserved.

A whole-repo AST sweep for @dataclass-decorated classes carrying
non-dunder methods now returns zero hits (verified below).

Also fixed one real regression risk found while sweeping: a
getattr(goal, "is_active", False) call site in recall_pipeline.py
silently defaulted to False once GoalVector.is_active became the free
function goal_vector_is_active, which would have disabled A3
goal-maintenance recall re-ranking with zero signal. grep for .attr
patterns alone does not catch getattr() defaults — caught by re-reading
every call site of every extracted method, not just the extraction itself.

Two more issues surfaced while re-verifying the mutation numbers below and
are fixed in this PR rather than left for a follow-up:

• tests_py/fuzz/test_corpus_replay.py imported replay_corpus.py via
  importlib.util.spec_from_file_location under the bare name
  "replay_corpus", not the package-qualified "fuzz.replay_corpus"
  mutmut's trampoline dispatch keys mutants by (mutmut run reported:
  "tests recorded trampoline hits but none match any mutant key"). Every
  mutant in that file was therefore invisible to mutation testing
  regardless of what the test asserted — the first table below originally
  (and wrongly) reported this file at 35/35/0; the real, re-verified
  number is 35/33/2 (below). Fixed via a normal from fuzz import replay_corpus import (behavior unchanged) plus two new negative-case
  tests for harness_consume's error paths and one for replay()'s
  exception-wrapping (also touched by this diff's call-site update).
• benchmarks/lib/__init__.py eagerly imported BenchmarkDB (→
  psycopg/psycopg_pool/pgvector) at package-init time, so importing
  the Postgres-independent benchmarks.lib.verification_report (one of
  the 26 extracted files) on a SQLite-only install raised
  ModuleNotFoundError before mutmut could run a single mutant against
  it. Every other consumer of BenchmarkDB in this package already
  defers that import inside a function for exactly this reason;
  __init__.py did not follow its own package's convention. Fixed via a
  PEP 562 module __getattr__, with a subprocess-isolated regression
  test (tests_py/benchmarks/test_lib_init_no_psycopg.py) that poisons
  sys.modules for psycopg/psycopg_pool/pgvector.

Mutation evidence (§12) — per-file, per-function, scoped runs

Per the issue's methodology note (a single combined multi-file mutmut run
gives unreliable "survived" verdicts under fork()+Pool() resource
contention — reproduced and discarded), every file below was run in
isolation via a scoped driver equivalent to scripts/mutation_check.sh
(temporarily repoints pyproject.toml's [tool.mutmut] only_mutate /
source_paths / pytest_add_cli_args_test_selection at the one file +
its test(s), runs mutmut run, restores pyproject.toml after). Two
files' isolated runs were themselves unreliable for reasons unrelated to
the combined-run bug — doctor_mcp.py's baseline-stats phase failed
outright (scripts/launcher.py missing from the mutants/ sandbox; fixed
by adding scripts to the scoped run's also_copy) and
replay_corpus.py's mutants were entirely uncounted (the module-
qualification mismatch described above) — both re-run clean after their
respective fixes; the table below reflects the corrected, re-verified
counts. 677 mutants generated across the 30 classes' extracted functions:

0 unexplained survivors. 4 functions generate 0 mutants (no mutable
comparison/arithmetic/literal AST node in a pure pass-through body) —
this is a mutmut operator-coverage limit, not a test gap; extraction
still moves them out of the @dataclass-skip blind spot for when a
future change adds real logic there.

replay() and main() (both pre-existing free functions in
fuzz/replay_corpus.py, not part of the Harness dataclass extraction,
but touched by this diff's call-site update from harness.consume() /
harness.inputs() to the new free functions) were swept for survivors as
a matter of course: replay()'s exception-wrapping path had one real gap
(no test made consume() itself raise), fixed with a new negative-case
test; main()'s 39 mutants are outside this diff's blast radius (its
body is unchanged, and it is already excluded from the pytest test
selection — a pre-existing CLI-only gap, not introduced or worsened
here).

Repo-wide dataclass-with-methods sweep (issue's own repro command)

Ran the issue's own reproduction script verbatim against the final tree:

$ python3 sweep_dataclass_methods.py   # issue #282's repro script, verbatim
(no output — zero @dataclass classes with non-dunder methods remain)

Test plan

• Full suite: pytest -q → 6868 passed, 5 skipped, 116 subtests
  passed, 0 failed (confirmed after the mutation re-verification and
  the two additional fixes above — mutation runs restore
  pyproject.toml and never touch source).
• ruff check + ruff format --check on all 77 touched files: clean.
• scripts/check_doc_claims.py: doc claims OK.
• scripts/generate_repo_badges.py --check: badges OK (4 checked) — no
  suite-size count reintroduced into prose/JSON docs per issue fix: eliminate the six-file test-count conflict class with a monotone floor (issue #293) #294.
• No conflict markers; .bestpractices.json parses.

§4 size caps — measured, not asserted

Every newly extracted function is within the 40-line cap (largest:
backpressure_pipeline_run at 33 lines). This PR does not claim a
repo-wide §4.1/§4.2 pass: several touched files (auto_curator.py,
recall_pipeline.py, doctor_mcp.py, procedural_memory.py, and others)
were already over the 300-line file cap on origin/main before this diff
(confirmed via git show origin/main:<file> | wc -l against each
flagged file, re-verified with a script diffing origin/main line counts
against the current tree) and contain pre-existing over-40-line functions
unrelated to the extracted methods (e.g. mine_skills, assess_conflict,
build_clusters) — untouched debt outside this diff's blast radius
(§14.3), not introduced or worsened here.

Two categories of genuine (small) growth, both within §10's explicit
Medium-stakes "≤20% flexibility on limits" band for rule 4:

• mcp_server/core/wiki_redirect.py crossed the file cap it was
  previously exactly at: 300 → 307 lines (+2.3%), from the extraction's
  rationale docstring (the same pattern used for every other class in
  this sweep) landing on a file that had zero headroom. Not split
  further — a dedicated file-size refactor of this module is outside
  this issue's scope (§15.1) and would mix an enabling refactor with this
  PR's mechanical extraction.
• A handful of already-over-cap files/functions grew by 2-10 lines each
  from mechanical call-site adaptation (x.method() → helper(x)
  sometimes needing an extra wrapped line): match_skills (+10),
  conflict_monitor_rerank (+2), apply_habituation (+2),
  run_distill_drain_cycle (+4), curate_wiki.py::handler (+2), and two
  test functions in test_headless_authoring_throttle.py /
  test_staging_resolve_sink.py (+2/+4). None introduce new logic; all
  verified against origin/main line counts per function.

Completion Ledger (§13.2)

Boy-scout (§14)

Defects seen and fixed in this PR (all in material this diff touches or
that this diff's own mutation-testing verification step executed
against):

1. getattr(goal, "is_active", False) stale-attribute regression in
   recall_pipeline.py (found while re-reading call sites of every
   extracted method).
2. tests_py/fuzz/test_corpus_replay.py's file-path import defeating
   mutmut's coverage tracking for the entire Harness extraction (found
   while producing the mutation-evidence table).
3. benchmarks/lib/__init__.py eagerly importing psycopg at
   package-init, breaking SQLite-only installs of one of the 26 extracted
   files (found the same way).
4. test_benchmark_db_resolves_with_psycopg_present (added for defect 3)
   itself assumed psycopg is always installed; CI's "Test (SQLite
   backend)" job installs without the Postgres extras and failed with
   exactly the ModuleNotFoundError the code under test is supposed to
   raise loudly — the test's premise, not the code, was wrong. Guarded
   with pytest.importorskip("psycopg") (caught by this PR's own CI run,
   fixed same-day before merge).

No other seen-and-unfixed defects; no bypass used (no skip flags, no
narrowed globs, no temp-dir dodges — each defect above was fixed at its
source and re-verified).

Scope note (§15)

The issue's own acceptance criteria suggested sequencing this as several
PRs (one per module family) given the size. Per the release-blocker
directive for this issue queue (no new issues, adjacent PRs only where
genuinely necessary), this PR completes the full 30-class sweep in one
diff rather than splitting further — the whole-repo AST sweep confirms
there is no remainder.