tests: miner.py — first coverage (validate_local_config + sha256_dir)

[roykollensvendsen]

Summary

First test coverage for miner.py (previously 0% covered). Two pure
filesystem functions tested with tmp_path — no mocking needed.

Functions covered

Function	Tests	Spec source	Outcome
validate_local_config(king_dir, challenger_dir)	7	docstring + sig + parallel inference from validator.validate_challenger_config	confirms-behavior + 1 resolved spec-ambiguity
sha256_dir(path)	5	sig + name + call sites + validator comment ref	confirms-behavior

12 tests, all green.

Coverage

validate_local_config — the miner-side equivalent of
validator.validate_challenger_config:

• Matching config + safetensors → returns None
• Missing king config → returns None (initial seed scenario)
• Arch mismatch → rejected
• Vocab mismatch → rejected
• Missing challenger config → rejected
• Missing safetensors → rejected
• .py file present → rejected (security)

sha256_dir — content hash of *.safetensors files. Critical
because the miner encodes this hash on-chain and the validator
recomputes it; any divergence silently invalidates every miner's
commit:

• Returns 64-char hex digest
• Deterministic
• Content-sensitive
• Non-safetensors files ignored
• Sort-invariant across filesystem ordering

Findings during testing

• Spec-ambiguity resolved: my initial happy-path test for
  validate_local_config only set up config.json. It went RED with
  "no .safetensors files in challenger" — the function also checks
  the directory contents. Reframed the helper (_make_valid_challenger)
  to include safetensors and added an explicit
  test_..._missing_safetensors_rejects test to lock in the
  discovered behavior. Also added test_..._python_file_rejects since
  the same security rule from validate_challenger_config applies.

Verification

.venv/bin/python -m pytest tests/test_miner_validation.py -v
# 12 passed
ruff check --config ../scripts/teutonic.ruff.toml tests/
# All checks passed!

Series context

PR 6 of foundation series. With this merged, foundation will have
68 tests across 11 functions/surfaces. Coverage now spans
validator.py, eval_torch.py, AND miner.py (no longer 0%).

PR 5 (new helpers from upstream
unarbos/teutonic#3)
remains the only deferred item — those symbols don't exist in the
current main yet.

[roykollensvendsen]

Self-review

Substantive (added in a951923)

• ✅ Multi-shard happy path for validate_local_config. Original
  happy-path test used a single model.safetensors. Real-world
  checkpoints are typically sharded (model-00001-of-N.safetensors).
  Added test confirms multi-file is accepted — catches a hypothetical
  regression where someone tightened the check to "exactly one file".

• ✅ Empty-directory test for sha256_dir. Documents defined
  behavior for the edge case (no .safetensors files): returns
  hashlib.sha256().hexdigest(). Real scenario: state cleared but
  directory not yet removed. Catches accidental crash-on-empty.

Verified

• All 14 tests passing locally + ruff clean
• Coverage spans both miner-side validation paths (config compare +
  content hash)
• TDD discovery preserved in commit history: 643c879 (initial,
  spec-ambiguity finding) + 904af2c (sha256_dir) + a951923
  (self-review edge cases)

Considered, not changed

• Subdirectory test for sha256_dir — would catch an accidental
  switch from glob to rglob. The sort-invariance test partially
  locks in the file-enumeration semantics. Marginal additional value.
• .pyc / hidden file pattern variations for the .py rejection
  rule. The impl probably uses a simple suffix check; testing edge
  cases would lock in arbitrary impl details.
• Performance/large-file behavior for sha256_dir. Correctness is
  the test target; perf is a different concern.

Series context

PR 6 of foundation series. With this merged, foundation will have
70 tests across 11 functions/surfaces. miner.py goes from 0% to
covered for both pure helpers; main() entry point remains untested
(intentionally — entry points typically aren't unit-tested).