…EVEL

Algorithm-level PARTIAL discharge for FALSIFY/INV-BPE-003 (round-trip
byte-equality on 10K held-out docs) per
`contracts/tokenizer-bpe-v1.yaml`. Also FALSIFY-SHIP-012 directly.

## What this binds

`verdict_from_roundtrip_scan(docs_scanned, roundtrip_failures)`
returns Pass iff:

1. `docs_scanned >= 10_000` (contract statistical floor)
2. `roundtrip_failures == 0` (zero-tolerance — non-injective
   tokenization corrupts loss target)
3. `roundtrip_failures <= docs_scanned` (counter sanity)

Pinned constant: `AC_BPE_INV_003_REQUIRED_DOCS = 10_000`.

## Why zero-tolerance

`decode(encode(nfc(text))) != nfc(text)` means the tokenizer is not
injective on the input domain. Even one failure means MODEL-2's
training loss is computed against tokens that don't reconstruct the
original text — silently corrupting the cross-entropy gradient on
that document. The contract falsifier ("Any non-zero diff bytes
fails") admits no tolerance band.

## Why 10K floor

The contract specifies a 10K-doc held-out corpus. Round-trip
failures cluster around rare Unicode patterns (emoji ZWJ sequences,
RTL combining marks, control sequences) that occur at frequencies
of ~1-in-10K to 1-in-100K. A scan over 1K docs would miss those
classes entirely; a scan over 100K would over-tax every smoke run.
Pinning 10K matches the contract's stat-power requirement.

## Why partition-violation check

`roundtrip_failures > docs_scanned` indicates counter corruption —
the failure set cannot exceed the total scanned. Catches a counter-
rollover bug or double-count regression.

## Five-Whys

1. Why bind INV-BPE-003 now? — Round-trip injectivity is the
   foundation of MODEL-2's training-target correctness; without a
   verdict pin, a regression in NFC handling or BPE merge-table
   loading silently corrupts every gradient.
2. Why a (u64, u64) pair, not the actual scanner? —
   Algorithm-level pin; the streaming round-trip scanner is
   FULL_DISCHARGE work for the corpus-tokenize PR.
3. Why pin 10K floor? — Statistical-power match to contract;
   smaller samples miss rare-Unicode classes.
4. Why partition-violation check? — Counter corruption mustn't
   silently pass.
5. Why 17 tests across 7 sections? — Mutation survey: provenance
   pin (×1), pass band (×4: floor, +1, CSN-Python, huge), zero-
   tolerance fail (×4 incl. one-in-million), sample-size fail (×3),
   partition violation (×2), boundary sweep (×8 each dimension), and
   zero-tolerance property (×4 sizes).

## Cross-reference

This verdict ALSO discharges FALSIFY-SHIP-012 (round-trip gate from
SHIP-TWO-001 spec §15.4) — the contract identifies the two as the
same gate. Two contract IDs share one PARTIAL discharge.

## Scope

PARTIAL_ALGORITHM_LEVEL only. Wiring this into the actual streaming
round-trip scanner that produces `(docs_scanned, roundtrip_failures)`
is FULL_DISCHARGE work for the corpus-tokenize implementation PR.

## Tests

17 unit tests, all green.

…ial-discharge

# Conflicts:
#	crates/aprender-core/src/format/mod.rs