feat(enhance): the degenerate exits get their own diagnostic channel — exit_kind + the incomplete_summary histogram (#615)

[gadievron]

Fixes #615.

What was wrong

The agentic enhance loop's four degenerate exits are conflated under one classification: "incomplete", but they differ in kind and remedy: three cheap model-behavior exits (end-turn-without-finish, finish truncated at max_tokens, no-tool-calls — confidence 0.3, remedied by retry/prompt work) vs one budget-exhaustion exit (MAX_ITERATIONS — confidence 0.2, remedied by raising the budget, not retrying). In the referenced run the split was 41 "Analysis incomplete" + 4 "Could not complete analysis within iteration limit" — a 4-unit class a consumer reading only classification cannot distinguish from the 41.

The fix

No fourth classification (that would break every == "incomplete" consumer — checkpoint.py, enhancer.py, the degenerate-exit tests). Instead:

• exit_kind on AgentResult — the four exit sites stamp their kinds; a completed analysis carries the empty string and serializes without the key (present-only — the completed-record byte-identity).
• The incomplete_summary histogram beside the existing error_summary: the enhancer's counting loop aggregates exit_kind (legacy unstamped rows fall back to "unstamped"; single-shot rows cannot reach the branch — their whitelist has no "incomplete"); EnhanceResult carries the histogram (present-only in to_dict); and both step-report writers — the scanner's pipeline summary and the standalone CLI's enhance summary — thread it, as FLAT siblings of the error_summary gate (a zero-error run with incompletes carries the histogram too — the pipeline round caught the CLI writer nesting it inside the error gate, which dropped it).
• The classification and every existing consumer are unchanged (pinned by the classification-unchanged test and the branch-adjacency source pin).

Evidence

• RED (pristine master, this file copied in): 7 failed + 1 passed (the 8-item set; the green row is the histogram mirror-test — noted).
• GREEN: 8/8 on the file — the four kinds + the completed negative (present-only), an incomplete serializing its kind, the classification-unchanged pin, the histogram census, the EnhanceResult carry + the present-only negative, the flat-writer pin on BOTH writers, and all FOUR real-exit receipts driving the actual agent loop (end-turn-without-finish, finish-truncated, no-tool-calls, max-iterations — real AgentResults, not hand-constructed; the harness already drove all four, the "two cheaply-simulable" claim was false). The full suite locally 4189 passed + 34 skipped + 2 failed — 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 de-balancing check (the intent trace)

serves: I-enh; de-balances-checked: the four == "incomplete" consumers are byte-identical (nothing reads exit_kind downstream except the histogram); the completed record serializes key-for-key identical to master (the present-only splat).

The adversarial record

The authoring rounds (the CLI second-writer catch, the real-exit receipts) + this PR-processing round (a combined 2d+DU seat that caught the CLI nesting bug — the histogram inside if result.error_summary:, a zero-error run with incompletes writing no histogram to the standalone enhance step report; a 4-seat panel — 3/4 approve with the CHANGELOG duplicate-heading fix, applied; two deep-refute seats): the dedent (both writers' flat placement pinned); the four real-exit receipts; the CHANGELOG entry (the repo's stated convention — the PR carried none) with the duplicated-heading artifact fixed; the branch-adjacency source pin (the histogram line inside the cls == "incomplete" branch — the placement-blind grep passed if the line moved outside and counted every unit's kind); the schemas comment corrected.

Notes for the reviewer

• The four EnhanceResult edits meet feat(analyze): the enhancement non-verdicts are visible, out of the prompt, counted, and migrated on re-enhancement (#611) #635's incomplete_count/total_units edits on merge — the union carries both (additive resolution); the post-merge invariant to pin: sum(incomplete_summary.values()) == incomplete_count (both increments sit in the same cls == "incomplete" branch — holds by construction at the merge).
• The context_enhancer's internal _summary_incomplete count and the resume path are kind-blind by design (they count the classification, not the kind) — out of this issue's stated scope.