Diagnose branch-aware duplicate frontiers

[iam-brain]

Question

Can the remaining ledger undercount come from deduplicating equal numeric token snapshots across branches after those branches have independently diverged?

The current ledger fingerprint is (last totals, cumulative totals) within a connected lineage family. That is strong evidence for copied history before a fork, but after branch divergence two independently billed turns can converge on the same numeric counters. Treating those as one observation would undercount without producing an unresolved-parent diagnostic.

Proposed slice

• model a branch frontier for each physical owner from the first observation not present in its parent/prefix
• deduplicate equal fingerprints freely inside the shared prefix
• after the frontier, require stronger copy evidence (stable event identity when available, or matching timestamp plus neighboring cumulative flow) before merging observations from different owners
• keep ambiguous post-frontier collisions contained rather than counting both blindly
• add aggregate diagnostics for shared-prefix duplicates, strong post-frontier duplicates, and ambiguous post-frontier collisions
• replay July 9–11 UTC and ordinary days before changing selected totals

Acceptance

• synthetic sibling tests prove copied prefixes count once and independently convergent suffixes count twice
• ambiguous collisions fail closed
• real replay reports how many tokens are affected before enabling the rule
• bounded memory remains family-local

This is a follow-up to #23 and should precede #20 if the diagnostic explains a material share of the residual.

[iam-brain]

Implemented and locally validated as shadow-only evidence in the new stacked PR. The frozen corpus found zero ambiguous post-frontier collisions and zero estimated suppressed tokens, so selected totals remain unchanged.