fix(llm-reach): provider-exception batches are counted — batches_failed, the shared units_not_reviewed, and the step's partial status (#541)

[gadievron]

Summary

The #386 counting covered the parse path only (on_batch_drop / _count_drop): a batch whose LLM call raised (empty completion, transport error) exited through the except-continue BEFORE the counter — so a step report could read status: success, batches_dropped: 0, units_not_reviewed: 0 for a pass that reviewed nothing (the receipt: 4 empty-completion batches on the run's first attempt, invisible to every surviving counter).

The fix:

• a distinct counter for the exception class (batches_failed — a different failure class with different remediation than the parse drops) + the same units_not_reviewed denominator so the coverage gap is the visible sum;
• the #285/#376 partial-status contract: the LLR step summary now carries error_count = batches_dropped + batches_failed, so all-batches-failed reads partial, never success/0/0 (the adversarial review's catch);
• both counters surface in the scanner's step summary.

Unchanged: the #538 truncation subclass (batches_truncated) is orthogonal; the absence-as-retry semantics (an exception batch still leaves no checkpoint records — it re-runs on resume); the #293 three-state summary arithmetic (the exception units stay in the incomplete bucket).

Closes #541.

Test plan

4 tests: the exception batch counts (batches_failed == 1, units_not_reviewed == 2, the parse counter untouched); a parse drop NOT counted as failed; a mixed pass reports both classes with units_not_reviewed as the sum; the step-status derivation — the counters in the summary make step_context read partial (never success/0/0).

Verification evidence

check	command	result
new + family	pytest tests/test_issue532_llr_resume.py -q	29 passed
full suite	pytest tests/ -q	3995 passed, 2 failed — both pre-existing (SDK pin drift)
static analysis	ruff / Semgrep	clean / 0
adversarial review	combined wave+refute	the refute's missed-item folded: the step-status gap (all-batches-failed reading success) closed via error_count; the #293 arithmetic and the auth-error class verified unchanged

[gadievron]

Gate-session receipt (proportionate round — astra; the counting-surface class)

Re-derived independently: the #532 family 29/29 (the 4 new tests included); the family (294 diagnostics + reachability + the #285 status contract) 48/48; full suite 3809/0 fresh-venv; ruff clean; the diff verified (the except-continue increment, the two stats keys, the scanner's error_count = batches_dropped + batches_failed wiring).

Astra: NOT-REFUTED — (a) the error_count sum is the ONE numeric source (the #285 contract's alternatives-not-additives shape); (b) the LLMAuthError re-raise-before-count correct (the designed abort, not a counted failure; the BaseException class escapes except Exception by construction); (c) no double-count mechanism established (the truncation continue and the provider-failure path are mutually exclusive per iteration — a raised call produces no completion for the truncation path to process); (d) the #293 three-state arithmetic unchanged (the exception units stay incomplete — the counters change observability, not classification).

CI green. Merge-ready.