Count parse failures in enhance telemetry

[gadievron]

Count parse failures in enhance telemetry

Branch: pr-2b-enhancer-telemetry (1 commit off master/d4caf8a) · Target: knostic/OpenAnt:master
Diff: 2 files, +42 (fix +4, test +38) · Not yet pushed — awaiting approval.

Problem

In the non-agentic enhancer (utilities/context_enhancer.py::ContextEnhancer.enhance_unit), when the LLM response can't be parsed, the else branch stores _get_default_context(error={"type":"parse_error"}) on the unit but does not increment self.stats["errors"] — while the exception branch already does. So the [Enhance] Errors: N summary under-reports parse failures.

This is telemetry-only: the unit still carries llm_context.error and is analyzed with default context (not dropped). Benign for coverage, wrong for the error tally.

Fix

Add self.stats["errors"] += 1 to the parse-failure else branch, mirroring the exception branch. One line, purely additive; no verdict/coverage behavior change.

Testing

• tests/test_enhancer_parse_failure_counted.py (2 tests): a parse failure increments errors (and is not counted as units_enhanced); a successful parse does not. Passes on this d4caf8a base.

Blast radius (checked)

• ContextEnhancer.stats["errors"] is the enhance-stage counter (surfaced in the [Enhance] Errors line). It is a different counter from the analysis-stage metrics["errors"] that core/reporter.py:442 reads ("units_analyzed": total_units - metrics.get("errors", 0), at the d4caf8a base) — so this fix does not change units_analyzed.

Provenance (direct re-derivation — NOT the KB grade-A record)

This fix stands on a direct re-derivation at the PR base d4caf8a: the parse-failure else branch stores
_get_default_context(error={"type":"parse_error"}) but does not self.stats["errors"] += 1, while the
adjacent except branch does — verified at d4caf8a, and the guard test fails on the pre-fix d4caf8a source.

Do NOT cite the KB enhancer-failed-context-no grade-A record as corroboration (per fable review): that
record describes a different symptom — "failed enhance stored with no error key" — which is the c81c0e1
state and is already fixed at d4caf8a (the else-branch there carries error={parse_error}). Its grade row's
stated basis ("byte-identical blob at c81c0e1") is also false — context_enhancer.py blobs differ (f90d29b≠e5365606).
So the grade-A record neither matches this PR's target nor transfers by identity. (Both are KB-provenance defects
flagged for feedback in KB-STUDY.md.) This PR's narrower counter-defect is real and residual at d4caf8a.

Rollback

Git-revert-clean: one commit, context_enhancer.py (+ its test), disjoint from all other changes. Reverting
restores the prior (under-counting) telemetry with zero coverage/verdict impact.