test(stm): consolidate the certificate proof tests around the golden proof

[hjeljeli32]

Content

This PR routes the verify-only tests in proof_system::halo2_snark::proof to the committed golden proof. Each of them previously built a full prover setup and generated its own proof, even though the assertion only ever needed proof bytes — a valid proof to tamper with, or any valid proof at all. Verification needs neither an SRS nor a proving key: SnarkProof::verify reads only the proof bytes plus the aggregate verification key, message and verifying key it is handed, and the verifier parameters are a 192-byte embedded constant.

Measured locally, --test-threads 1: the slow tier goes from 33.9 s to 15.3 s warm (-55 %) and from 65.8 s to 47.1 s cold (-28 %), 16 tests becoming 15. Fast tier unchanged at 511 passing. Per-item numbers are in #3463.

Changes

• Verify negatives (proof.rs): verify_fails_with_random_bytes_or_wrong_number_bytes, verify_fails_with_wrong_message and verify_fails_with_wrong_circuit_verification_key move into mod golden and run against golden_proof(). Four of the routed tests previously ran at m=100, k=5; verify and prepare_and_check never read the proof's own params or merkle_tree_depth, so the golden proof's configuration serves them without changing what they assert.
• valid_proof_verifies deleted: once routed its body is identical to golden_conversion, which is itself unmodified.
• Wrong verifying key: obtained by decoding NON_RECURSIVE_CIRCUIT_VERIFICATION_KEY_FOR_PRODUCTION instead of building a second prover. MidnightVK is self-describing, so this needs no SRS and no keygen, and the assertion strengthens to "a test proof must not verify under the key that ships".
• Prepare-and-check and codec tests: the three prepare_and_check tests and snark_proof_to_from_bytes routed the same way, every assertion kept.
• Fresh-proof coverage: produces_valid_snark_proof gains a prepare_and_check assertion on the proof it already generates, so that method still runs against a freshly produced proof rather than only committed bytes. Costs 0.03 s.
• Untouched: valid_snark_proof_with_different_path_length and non_deterministic_proofs_verify, the two tests that fail during aggregation before proving, different_parameters_prove_and_verify_fails, and the two pre-existing golden tests are byte-identical to main. produces_valid_snark_proof is the only surviving test that changed, by the assertion above.
• Not done, the cached forged proof: implemented and measured, then reverted. Locally it measures +0.65 s on a cold cache against -0.47 s on a warm one, and the content-keyed cache directory is empty on every CI run, so the cold path is the one CI takes. Left for @jpraynaud's decision, with the numbers in a comment on Consolidate the halo2_snark slow proof tests around shared and golden proofs #3463.

Pre-submit checklist

• Branch
  • Tests are provided (if possible)
  • Crates versions are updated (if relevant)
  • CHANGELOG file is updated (if relevant)
  • Commit sequence broadly makes sense
  • Key commits have useful messages
• PR
  • All check jobs of the CI have succeeded
  • Self-reviewed the diff
  • Useful pull request description
  • Reviewer requested
• Documentation
  • No new TODOs introduced

Issue(s)

Closes #3463

[github-actions]

Test Results

     5 files  ± 0     209 suites  ±0   1h 59m 29s ⏱️ + 36m 59s
 3 449 tests +51   3 449 ✅ +51  0 💤 ±0  0 ❌ ±0 
11 312 runs  +54  11 312 ✅ +54  0 💤 ±0  0 ❌ ±0 

Results for commit 7d066b7. ± Comparison against base commit 4ec7b25.

♻️ This comment has been updated with latest results.

[damrobi]

LGTM 👍
I just think that you can probably move some of those test function outside of the slow module, maybe in a golden module outside of slow?