audit: classify FFI binding unsafe blocks as legitimate (PA001/PA007, 129 entries) (#67)

## Summary

`panic-attack assail` reports 88 `UnsafeCode` (PA001) + 41 `UnsafeFFI`
(PA007) Critical/High findings across this repo's binding-language
directories. All sit at the C-ABI boundary: every
\`bindings/<lang>/src/\` directory is a *thin wrapper* over the C ABI
exposed by \`libproven\` via the Zig FFI bridge. **No logic is
reimplemented in any binding** — every function calls through to the
formally-verified Idris 2 implementation.

This PR adds the user-classification registry so the next assail pass
suppresses these 129 findings, leaving any **new** unsafe code outside
the FFI roots visible.

## What changes

- \`audits/assail-classifications.a2ml\` — 129 \`(classification ...)\`
entries, all \`classification="legitimate-ffi"\`.
- \`audits/audit-ffi-bindings-2026-05-26.md\` — per-binding rationale.

## Scope

The classification is **scoped to FFI-boundary roots**:

- \`bindings/*/src/\`
- \`ffi/*/src/\`
- \`domain-specific/*/ffi/*/src/\`

Any unsafe block outside those roots remains visible. The audit doc
enumerates exactly which binding languages are covered (Ada,
AffineScript, C, Elixir, ephapax-{affine,linear}, Go, Haskell, Rust,
SML, VCL, Zig, BEAM NIF, http-Zig).

## Why classification, not code fix

You can't \"fix\" FFI unsafe — it's structural. The C ABI requires
unsafe at the boundary in every host language. The reviewable record is
the registry + audit pair: any *new* unsafe block introduced inside the
FFI roots needs a companion classification entry, which is a separate
visible edit.

## Anti-gameability

Same pattern as \`hyperpolymath/007\` and \`hyperpolymath/echidna\`.
Registry is a separate file from any source under scan; suppression
cannot be added one-sidedly without a companion audit-doc edit.

## Verification

Locally on this branch: \`panic-attack assail . --headless\` reports:

| Category | Active (Critical/High) | Suppressed |
|---|---|---|
| UnsafeCode (PA001) | 0 | 88 |
| UnsafeFFI (PA007) | 0 | 41 |

(The 1 remaining UnsafeCode and 8 UnsafeFFI are Medium severity, out of
scope this PR.)

## Out of scope this PR (separate follow-ups)

| Category | Count | Notes |
|---|---|---|
| CommandInjection (PA003) | 15 | Mostly tooling/build scripts |
| HardcodedSecret (PA009) | 6 | Mostly test fixtures |
| DynamicCodeExecution | 5 | Separate triage |
| SupplyChain (PA023) | 2 | Version-pinning judgement |
| UnsafeTypeCoercion / UncheckedAllocation / PanicPath | 5 | Spot triage
|

## Test plan

- [ ] CI green (hypatia-scan, secret-scanner, codeql, governance,
scorecard)
- [ ] Post-merge: \`panic-attack assail .\` shows 0 active Critical/High
UnsafeCode + UnsafeFFI
- [ ] No findings outside FFI roots reduced

Refs hyperpolymath/panic-attack#32 (estate sweep tracker).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: hyperpolymath