fixup(factors): address strict_bench code review findings

Independent adversarial review (run before upstream maintainer
review) surfaced 10 issues in run_bench_strict. This commit fixes
the high- and medium-impact ones in-place and adds 14 regression
tests. The strict bench's design contract is unchanged; the changes
are correctness and back-compat-only.

Fixes
-----

A1. OOS train/test boundary date no longer double-counts.
    .loc[:t] and .loc[t:] are both label-inclusive in pandas, so the
    split date previously appeared in both buckets. Replaced with
    explicit comparisons:
        train = alpha_full[alpha_full.index <= oos_ts]
        test  = alpha_full[alpha_full.index >  oos_ts]
    Also added ic_count_train / ic_count_test per row so
    categorise_strict can later enforce per-bucket min_ic_count.

A2. compute_random_ic_series now uses an inner join across seeds.
    pd.concat(axis=1, join='inner') ensures every retained date is
    the mean of *all* available seeds — not a hodgepodge of 1-seed
    and 5-seed averages on dates where some seeds dropped out due to
    the _MIN_VALID_PER_DATE=5 guard.

A3. on_progress callback is now exception-safe in every branch.
    Refactored to a single _fire_progress() helper invoked from both
    the empty-IC continue path and the normal end-of-loop path.

A4. _shuffle_within_rows pins ±inf in place like NaN.
    Switched the mask from ~np.isnan() to np.isfinite() so an
    inf/-inf cell stays at its original position. Defensive against
    third-party zoos that bypass _validate_output.

A5. OOS sign-flip is now categorised reversed_strict, not train_only.
    Bucket order: t_full >= thr AND t_test <= -thr → reversed_strict
    (most diagnostic failure); t_full >= thr AND t_test in noise band
    → train_only (benign decay).

A6. Per-bucket ic_count fields surfaced on each row.

A7. Sorting uses unrounded _ir_raw / _alpha_t_full_raw / _ic_mean_raw
    helper keys to keep top-N stable across runs.

C1. Wire schema regression — strict result now carries legacy aliases
    alive/reversed/dead/by_theme alongside the strict-specific keys.
    Existing dashboards keep rendering without code changes:
        alive    = confirmed_alive
        reversed = reversed_strict
        dead     = noise + train_only
    by_theme is built by a strict-aware variant that emits both the
    new four-way and the legacy three-way counts per theme.

C2. _slim payload re-adds formula_latex so the wiki / dashboard top-N
    cards keep showing the formula column.

C5. Error envelope is now schema-complete from the start. Every error
    path (empty zoo, bad universe, bad forward-returns, bad oos_split)
    returns a dict with zeroed counters, empty lists, and the rail
    metadata intact — downstream consumers can depend on every key
    being present regardless of status.

C6. n_random_seeds=0 is clamped to 1, AND the effective value is
    persisted to entry['n_random_seeds'] so the wire response doesn't
    lie about the seed count.

Internal sort-helper keys (_ir_raw etc) are stripped from public_rows
before they reach the wire payload — only _category is retained so
external consumers can read the bucket label.

Tests
-----

14 new regression tests in test_bench_strict.py — total now 33
strict + 988 existing = full agent/tests/factors/ suite is 1002
passed, 1 skipped, 24.25 s. Zero regression in the existing factor
tests.

The new tests:
- test_oos_train_test_split_does_not_double_count_boundary (A1)
- test_compute_random_ic_series_inner_joins_seed_dates (A2)
- test_run_bench_strict_on_progress_exception_is_caught (A3)
- test_shuffle_handles_inf_like_nan (A4)
- test_categorise_oos_sign_flip_is_reversed_strict_not_train_only (A5)
- test_categorise_oos_decay_to_noise_band_is_train_only (A5 companion)
- test_run_bench_strict_emits_legacy_alive_dead_reversed_keys (C1)
- test_run_bench_strict_legacy_alive_equals_confirmed_alive (C1)
- test_run_bench_strict_top_lists_include_formula_latex (C2)
- test_run_bench_strict_empty_zoo_returns_schema_with_counters (C5)
- test_run_bench_strict_n_random_seeds_zero_is_clamped (C6)
- test_run_bench_strict_rows_drop_underscore_prefixed_sort_keys (sort key
  hygiene)
- test_run_bench_strict_catches_planted_alive_signal (closes the
  'integration tests cheat' finding)
- test_run_bench_strict_catches_planted_reversed_signal (same)

Signed-off-by: Soli22de <177382421+Soli22de@users.noreply.github.com>

Author: Soli22de