Feature/Fix: Diagnose Codex branch-frontier collisions

[iam-brain]

Summary

• add a bounded, family-local shadow analyzer that separates copied prefixes from equal numeric token snapshots after branch divergence
• require resolved parent edges, prefix-compatible physical copies, independent divergence witnesses, and strong neighboring-flow evidence before classifying suffix collisions
• emit aggregate-only diagnostics with overflow, cancellation, UTC, and one-million-observation family guards
• keep the analyzer structurally disconnected from accounting rows, caches, and selector inputs

Real-world result

This theory also does not explain the remaining July 9 undercount. The frozen corpus resolved 365 parent edges and identified 314 shared-prefix fingerprints, but found zero ambiguous post-frontier collisions and zero estimated suppressed tokens on every day.

Selected totals remain unchanged. The negative result points the remaining July 9 gap toward discovery, parsing, or reference semantics rather than numeric fingerprint convergence after forks.

Tests

• focused frontier/shadow/local-validation suites (8 tests passed before review; focused frontier suite passed after review fixes)
• make check
• opt-in frozen replay over 819,779 observations
• make test (649 selections in 55 groups; 55 passed, 0 failures, 0 retries)

Closes #39. Stacked on #41.