- Addresses TODO #17755 by resolving the bounds-related issues and
documenting the current leaf-read semantics in
`content_addressed_append_only_tree.hpp`:
  - `leaf_index >= max_size_` → failure (out of tree range)
- `leaf_index < max_size_` but leaf not written → return 0 with `success
= true`
- (Historical reads at a block) `leaf_index >= blockData.size` → failure
(out of block range)
- (Historical reads at a block) `leaf_index < blockData.size` but leaf
not written → return 0 with `success = true`
- Update audit headers

## Summary
Hardens barretenberg verifier code paths and API boundaries based on
security review findings.
### Verifier robustness
- Replace `BB_ASSERT` with context-appropriate failures in verifier code
paths — `throw_or_abort` for parsing/structural invariants, return false
for native verification failures, circuit constraints for recursive
verifiers
- Move IPA SRS size check before circuit construction in
`full_verify_recursive` (fail early on misconfiguration)
- Remove redundant witness-value comparison in recursive IPA verifier
(`assert_equal` is the real enforcement)
- Guard unsigned underflow in batched translator verifier's
num_public_inputs derivation
### Input validation
- 2513: Reject non-canonical field encodings in `MsgPack`
deserialization (`value >= modulus` now throws instead of silently
reducing)
- 2501: Reject trailing data in `ChonkProof` `MsgPack` deserialization
(consistent with hardened `PrivateExecutionStepRaw` loaders)
### API state machine fixes
- 2504: `ChonkStart`: clear stale `loaded_circuit_*` state when starting
a new session, preventing silent reuse of circuits from a prior session
- 2509: `ChonkAccumulate`: clear loaded state immediately after move,
preventing use-after-move on retry if a later step throws
### Debug
- 2479: Fix off-by-one in `has_last_app_been_accumulated` flag
(`num_circuits - 4` → `num_circuits - 3`), debug-only
                  
### Tests
                  
- 2513: Regression test for non-canonical `MsgPack` field encoding
rejection (all 6 field types)
- 2493: Regression test for translator `NonNativeFieldRelation`
accumulator alias attack

---------

Co-authored-by: iakovenkos <sergey.s.yakovenko@gmail.com>

…nical field encoding (#22333)

## Summary

Fixes CI failure in `merge-train/barretenberg` caused by the interaction
of #22233 (Fr::from_u64 big-endian encoding) and #22311 (strict
non-canonical field validation).

The `corruptProofFields()` test helper was flipping high-order bytes
(indices 0,1) of field elements, creating values exceeding the BN254 Fr
modulus. The strict validation now rejects these at deserialization time
with `non-canonical encoding (value >= modulus)` before they reach the
verifier.

**Fix**: Flip low-order bytes (indices 30,31) instead — still corrupts
the proof but keeps field values canonical.

## Test plan
- [x] `batch_verifier.test.ts` — all 4 tests pass locally (previously 1
failing)

ClaudeBox log: https://claudebox.work/s/e5146dc92fea29dd?run=1

## Summary

Fixes WASM CI failure for `PrimeFieldTest/0.MsgpackRejectsNonCanonical`
in `ecc_tests`.

WASM builds define `BB_NO_EXCEPTIONS`, which makes `throw_or_abort()`
call `abort()` instead of throwing `std::runtime_error`. The test uses
`EXPECT_THROW` which cannot catch an abort — the WASM process crashes
with `wasm trap: unreachable`.

Wraps the test with `#ifndef BB_NO_EXCEPTIONS` and adds a `GTEST_SKIP()`
fallback, matching the existing pattern in `c_bind_exception.test.cpp`.

## Test plan
- [x] `ecc_tests --gtest_filter=*MsgpackRejectsNonCanonical*` — all 6
field types pass on native build

ClaudeBox log: https://claudebox.work/s/e5146dc92fea29dd?run=3