Audit remediation: host build, public API + Law I model binding, BNN/MLP correctness, engine tests

[anjaustin]

Summary

This PR is a head-to-toe audit and remediation pass on Minerva. It repairs the
documented host build, makes the public API match the docs, closes a Law I
verification gap, fixes correctness/memory-safety bugs across the BNN and MLP
paths, and adds the engine's first end-to-end test coverage. History is kept
as 9 small, atomic, independently-reviewable commits — happy to split into
themed PRs or squash if you prefer.

All work was verified first-hand: warning-free build, ctest green, and every
engine architecture checked against an independent integer reference under
AddressSanitizer + UndefinedBehaviorSanitizer. Each bug fix ships with a test
that fails on the old code and passes on the new.

What's fixed (commit by commit)

1. Fix host CMake build — the documented cmake && make && ctest flow
   never compiled: include paths omitted src/security (so mnv_engine.c
   failed on #include "mnv_blake2s.h") and MNV_SOURCES was missing
   mnv_lut.c / mnv_outauth.c. Also fixes the one -Wall -Wextra warning
   that surfaced once the engine TUs compiled.

2. Repair public API + bind verified model to context — mnv_run_with_model
   (used by the README quick-start and the example) was never declared in
   minerva.h, so consumers got a compile error; mnv_run() was a stub
   returning MNV_ERR_CONFIG; mnv_verify_output() was declared but never
   defined. Now mnv_init() binds the verified model into the context and
   mnv_run()/mnv_verify_output() operate on it. This also closes a Law I
   gap: mnv_run_with_model() previously accepted any model pointer at run
   time while only checking ctx->verified (set for whatever model mnv_init
   saw), so mnv_init(ctx, trusted) + mnv_run_with_model(ctx, other, …) ran
   unverified weights. Inference now rejects any model but the bound one.

3. Add end-to-end engine tests — the suite previously exercised only crypto
   primitives; mnv_init and the forward passes had zero coverage. New tests
   build a real ChaCha20+BLAKE2s model in memory and check output against an
   independent reference, plus tamper/null/bad-ABI paths. Driven by
   tests/host/run_engine_tests.sh (one build per arch, ASan+UBSan), wired
   into ctest.

4. Fix multi-layer BNN — (a) mnv_config.h auto-defined MNV_QUANT_Q8
   whenever absent, silently overriding MNV_QUANT_BINARY/Q4/Q15;
   (b) the BNN forward decrypted from offset 0 every layer (keystream advanced,
   ciphertext pointer did not → garbage past layer 0); (c) it never consumed
   the per-layer bias bytes the compiler emits. Now tracks the offset, decrypts
   and applies biases, and binary quant is actually selectable.

5. Size per-layer buffers to the widest layer — the activation buffers,
   weight scratch, and bias scratch all assumed layer 0 was widest; a wider
   hidden/output layer overflowed them (ASan: stack-buffer-overflow). Sized to
   the true max across layers, with a tightened static_assert.

6. Fix BNN dot product for sub-byte widths — when in_sz % 8 != 0 the
   byte-popcount path mis-aligned per-neuron weights and counted padding bits
   as agreements. Replaced with a branchless, bit-addressed XNOR dot
   (constant-time, any alignment).

7. Fix confidence-check signedness — it compared (uint8_t)max_val, so a
   negative max logit (e.g. -1 → 255) passed any threshold. Now a signed
   comparison, with its own MNV_ENABLE_CONFIDENCE_CHECK flag.

8. Add LICENSE + .gitignore — the README and citation claim MIT and
   reference a LICENSE that didn't exist; .gitignore keeps device keys /
   generated models from being committed by accident. (See note below on the
   copyright line.)

9. Docs + metadata sweep — version drift (header said 1.1.0, citation
   1.2.0 → 1.3.0), a false "keystream precompute" claim in the mnv_init doc,
   the mac "over plaintext" comment (it's over ciphertext / encrypt-then-MAC),
   blinded-LUT wording, reentrancy notes on the file-scope scratch buffers, a
   v1.3 changelog, and threat-model corrections (blinded LUT + output auth
   shipped; confidence default).

How to verify

cmake -S . -B build -DMNV_TARGET=host && cmake --build build && ctest --test-dir build -V
bash tests/host/run_engine_tests.sh   # all arches, ASan+UBSan

Things that are your call (not asserted unilaterally)

• LICENSE copyright line. I added the MIT text your README already promises,
  but used a placeholder holder (MINERVA project contributors). Tell me the
  name you want, or I'll drop the LICENSE from this PR and raise it as an issue
  — entirely your decision.
• Two design choices you may want to weigh in on: caching the model in
  ctx (vs. keeping the explicit-model entry points), and giving the BNN
  biases. Both are isolated commits; I can rework or drop either.

Disclosure

This audit and the fixes were done with AI assistance (Claude); the
Co-Authored-By trailers reflect that. Every change was built, tested, and
sanitizer-checked before submission.