…rate eigenvalue

P1 fixes (real / experiment-impacting bugs):

  * mad_scc code task: `_vote`/`_group_for_diag` now uses body-stripped BLEU
    for clustering (matches scc_components.voting._code_bleu_plurality).
    Previously full-code BLEU was used, so trigger (count_first_plurality,
    body-stripped) and vote (_group_for_diag, full-code) could disagree on
    cluster membership when agents used same logic with different
    docstrings. Verified: two `def add(a,b)` with different docstrings
    now correctly cluster as one group.

  * inference.py --sequential: process_sample() was called with 4 args but
    the function signature requires 5 (lock missing). Sequential mode
    crashed with TypeError immediately. Added the missing lock arg.

  * pc1_contributions: when all embeddings are identical, S_c ≈ 0 and eigh
    picks an arbitrary orthonormal basis. The leading "eigenvector" carries
    no information, but softmax(|v_1|) would give one random agent ~100%
    weight. Now detects raw_lam1 < 1e-10 and returns uniform 1/N.
    Verified: 5 identical embs now yield [0.2]*5 instead of [~1, ~0, ~0, ~0, ~0].

P2 fixes (consistency):

  * mad_scc/soo_scc pool-size handling unified to `>=`. mad_scc previously
    required strict equality (rejected oversized pools), soo_scc accepted
    and sliced. Now both accept `>=` and use the first N entries.

  * task_typing._MCQ_SOURCES now includes aqua-rat, medmcqa, medqa.
    Previously these relied on the option-pattern fallback in
    _query_looks_like_mcq, which has false-positive risk for math
    problems mentioning "(A)" patterns.

Round 3 P3+ findings deferred per "只修 P1 三条 + P2 顺手两条" decision:
  R1-MASWorks#2/MASWorks#3/MASWorks#4/MASWorks#7/MASWorks#8, R2-M2/M3/M6/M7/M8/M9/M10, R3-C2/C3/C5/C6.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>