fix(enhance): a completed paid classification survives a post-result assembly failure — the finish validator enforces include_functions element types (#614)

[gadievron]

Fixes #614.

What was wrong

The agentic enhance loop stored a completed, paid classification and then overwrote it with an error dict when post-result assembly failed:

• agent.py assigns the finished, billed classification to the unit's agent_context.
• The assembly loop then dereferenced the model's raw include_functions finish input — and the finish validator checked field presence only, never element types. A bare-string item raised AttributeError: 'str' object has no attribute 'get'.
• The handler then caught the raise and replaced the stored context with the error dict — destroying the completed classification and its metadata.

The referenced run's receipt: all 12 raise-class enhance errors are exactly this AttributeError — the paid classification existed, was stored, and was then silently discarded.

The fix (three layers)

1. The validator: include_functions must be a list of objects with string ids — enforced at the finish schema boundary (the model self-corrects on rejection). (The repo's own degenerate-exit test fixture carried the bare-string shape — presence-only validation had let it through; the fixture is corrected to the documented contract.)
2. Preserve on failure: agent.py's assembly block is wrapped — a raise marks assembly_error inside the stored context (the classification, reasoning, confidence, and recorded usage all survive; the additional code is simply not inlined). Never overwrite paid work with an error dict. The origin metadata computes into locals before the commit (a metadata raise no longer leaves inlined code with deps_inlined unset). The wrap does NOT re-raise — the preservation is delivered by the marker alone.
3. The handler distinguishes the two error shapes: an assembly_error-marked context is preserved and re-classified from it (the completed verdict exists — the summary/stats keying stays consistent); a genuine agent-loop failure (an LLM/parse raise with no completed context) still takes the error dict. Reachability, disclosed: the wrap does not re-raise, so the branch is defense-in-depth for a future raise-after-store.

The agentic stats no longer count units_with_context/functions_added for assembly-error units (delivery that failed); the verdict bucket still counts them (a completed analysis whose inlining failed is still a verdict).

Evidence

• RED (pristine master, this file copied in): 4 failed + 2 passed (the 6-item set — the two validator rows, the preserve, the handler; the valid-elements and the then-vacuous control are the green rows).
• GREEN: 9/9 on the file (pytest -q tests/test_issue614_finish_payload.py → 9 passed) — the bare-string / non-list / non-string-id validator rows; the preserve receipt driving the real enhance_unit_with_agent with a poisoned index (the completed neutral classification + reasoning survive, error is None, the marker carries the exception class); the defense-in-depth receipt — _finish monkeypatched to the regressed presence-only shape, the historical bare-string payload driven through the real entry point (the wrap catches at the historical AttributeError site; the classification survives; the handler never fires); the behavioral handler receipt — the real enhance_dataset_agentic closure driven with a store-then-raise fake (the context preserved, the classification re-read from it, never the error dict); the stats-gate receipt (an assembly-error unit: no units_with_context claim; the verdict still bucketed). The full suite locally: pytest -q → 2 failed, 4208 passed, 34 skipped — the 2 are test_llm_sdk_contract_floor SDK-pin host drift, identical on master; CI is the authoritative suite. Ruff clean; Semgrep 0 findings.
• The fixture correction re-routes three existing tests to their intended paths (the bare string previously never reached the assembly — analyze_unit does not run it; with the new validator the old fixture would have broken test_complete_finish_still_accepted and The agentic enhance give-up classes are conflated under one 'incomplete' classification — expose exit_kind diagnostics (the iteration-limit budget exhaustion differs in kind and remedy) #615's finish_truncated receipt).

The de-balancing check (the intent trace)

serves: I-enh; de-balances-checked: the genuine-failure error dict is byte-identical to master (the preserve only widens); every == "incomplete" consumer unchanged (#615's exit_kind stamping untouched — the wrap is disjoint).

Notes for the reviewer

• The error_summary grouping follow-up from the issue stands: the preserved-context path reduces the error cohort to genuine failures, where a finer grouping by exception_class becomes meaningful.
• The assembly's remaining raise sites (a non-dict index return, a non-str language) are all inside the wrap — they preserve the classification by construction; the validator closes the front door for the historical shape.