PR #457 R7: label first-period boundary as explicit Deviation (P1)

R7 surfaced a real P1: the REGISTRY presented the always-treated remap
condition `first_treat <= min(time)` as "per paper footnote 11", but
the paper's strict rule is `t_i < 1` (units treated *before* the first
observable period). The inclusive `<= min(time)` rule additionally
folds `first_treat == min(time)` cohorts into U — that's a library
boundary convention, not a paper-faithful rule. The test class
docstring already called this out, but the authoritative REGISTRY
contract did not, which read as an undocumented methodology deviation
on PRs that promote BaconDecomposition to **Complete**.

Resolution: added a new explicit `**Deviation (first-period boundary
extension on always-treated remap)**` block to REGISTRY's Bacon section
that:
- Names the paper's strict `t_i < 1` rule
- States the library's inclusive `<= min(time)` rule
- Explains the rationale (`first_treat == min(time)` cohorts have no
  untreated cell in-panel)
- Notes R does NOT apply this fold (it keeps such cohorts in their own
  bucket and emits `Later vs Always Treated`)
- Notes the rules coincide when `min(time) > 1`

Mirrored in:
- REGISTRY Assumption checks bullet (line 2619): now points at the
  new Deviation block
- REGISTRY `**Note (always-treated remap)**` (line 2684): qualifies
  the "per paper footnote 11" claim
- METHODOLOGY_REVIEW.md Deviations block: re-titled to include
  paper deviations, added the boundary entry as item 1
- `bacon_decompose()` docstring (`bacon.py:467-487`): explicit
  boundary-extension paragraph with REGISTRY pointer
- CHANGELOG PR-457 Added entry: explicit boundary-deviation callout

Also fixes R7 P3: CHANGELOG PR-B test count "~24 tests" updated to
acknowledge the post-release 34-test count after R-parity-goldens
expansion.

Tests: 34/34 pass.

Author: igerber