Late-perceptual report polish (calibration, sensitivity, disagreements) + ADR-0003

## Parent

Plan: `docs/superpowers/plans/2026-07-20-causal46-late-perceptual-significance.md` (Step 6). Depends only on the parquet from the core notebook — runs parallel to the gate swap.

## What to build

Enrich the `population_summary.pdf` emitted by `late_perceptual_significance.py` into the full report, and record the decision as an ADR.

Report panels:
- Per-window run-length histogram (the mixed short+diffuse regime that motivates TFCE).
- Param-sensitivity panel: TFCE with E in {0.5, 1.0}, H in {1, 2} — reported, not tuned (D3).
- Integral-vs-TFCE agreement (knob-free robustness column vs the gate statistic).
- 2x2 calibration table: TFCE-pass x manual `behav @late` over the 187, with sensitivity/specificity framed explicitly as calibration, NOT validation (D6).
- Disagreement cells listed **by name**: TFCE-pass and not-manual (candidate manual misses); manual and not-TFCE-pass (the 19 fallback + short cells that don't survive correction) — for eyeball follow-up only.

ADR-0003:
- Draft `docs/adr/0003-*.md` mirroring ADR-0001's structure: what moved (manual `behav @late` gate -> automated TFCE), why, consequences, and the manual set retained as calibration only.

## Acceptance criteria

- [ ] `population_summary.pdf` includes run-length histogram, param-sensitivity panel, integral-vs-TFCE agreement, and the 2x2 calibration table.
- [ ] Named disagreement cells listed in both directions.
- [ ] Calibration framed as calibration (manual labels never tune the gate or operating point).
- [ ] ADR-0003 drafted, mirroring ADR-0001 structure; a follow-up per the plan's header note.

## Blocked by

- #10 (produces `site_results.parquet` + the base `population_summary.pdf`)

Start with the plan in docs/superpowers/plans/2026-07-20-causal46-late-perceptual-significance.md and the output from #10, then inspect late_perceptual_significance.py and the existing population_summary.pdf. Use site_results.parquet as the input and mirror ADR-0001 when drafting docs/adr/0003-*.md. Done means the four requested report panels, named disagreements, explicit calibration framing, and the ADR are present.

Written by the indexing model from the issue text.