Tracking: causal46 late within-completion perceptual significance (TFCE gate)

Tracking issue for the late-window perceptual significance work — replacing the manual `behav @late` gate with an automated per-cell TFCE permutation test.

Plan: `docs/superpowers/plans/2026-07-20-causal46-late-perceptual-significance.md`

## Tickets

- [x] #8 — Tied-cell audit: confirm the persisted b4 null is usable (no blockers) — DONE 2026-07-20, 0/187 tied, drop-and-document confirmed
- [x] #9 — TFCE + max-null helpers in `_windows.py` (no blockers) — DONE 2026-07-20, tfce_enhance/max_tfce_null/assert_coherent_null_replicates + 11 unit tests
- [ ] #10 — `late_perceptual_significance.py` notebook + Snakefile rule → `site_results.parquet` + headline (blocked by #8, #9)
- [ ] #11 — Swap late-window entry gate to the TFCE set + wire Snakefile dependency (blocked by #10)
- [ ] #12 — Report polish (calibration, sensitivity, disagreements) + ADR-0003 (blocked by #10)

Dependency shape: `#8, #9 → #10 → {#11, #12}`.

Start with docs/superpowers/plans/2026-07-20-causal46-late-perceptual-significance.md and review the dependency chain from #8 and #9 to #10–#12. Inspect late_perceptual_significance.py and the Snakefile rule, then verify that the work produces site_results.parquet and the headline. Done means the TFCE gate is wired, dependencies are updated, and report polish plus ADR-0003 are complete.

Written by the indexing model from the issue text.