test(cli): cover partial-batch --json, --apply-same and model-not-found paths (#468)

[jasonssdev]

Closes items 1-3 of the partial-batch follow-up list. Refs #468, which stays open for items 4-5.

What was unproved

adjudicate's docstring states that every output mode (report, --json, --apply, --apply-same) processes batch.results exactly as a complete run before reporting the failure. Only the default report and --apply had partial-batch tests. And curate's OllamaModelNotFound re-raise arm was exercised only through a sequencer-level fake stage that raises directly — never through a real partial AdjudicationBatch.

The docstring is left intact. These tests prove the claim rather than weaken it.

The three tests

Test	Covers
test_adjudicate_json_partial_batch_emits_completed_verdicts_then_exits_one	--json: stdout parses to exactly the completed verdict, stderr carries the completed-of-total line, exit 1
test_adjudicate_apply_same_partial_batch_confirms_completed_count_only	--apply-same: Total: and the confirm gate use the COMPLETED count; the merge commits; exit 1 follows
test_identity_partial_batch_model_not_found_still_walks_then_skips_later_stages	OllamaModelNotFound through _identity_run's returned-batch split, with a merge already applied

The --json test is the RETURNED-batch path; the pre-existing unavailable test covers the RAISE path, where stdout stays empty. That pairing is what makes a partial array distinguishable from the pre-#441 behavior.

What item 4 looks like in practice

The --apply-same test pins current behavior rather than proposing a fix. Two groups queued, one adjudicated SAME: the operator is shown Total: 1, types 1, the merge commits — and only then does stderr reveal that a second group was queued and failed. The preview is accurate for what gets merged, but nothing at confirm time discloses the partial-ness. Whether it should is still open, and it now spans three surfaces (see below).

Verification

• uv run pytest -q → 4035 passed, 1 skipped (the known darwin-gated test).
• uv run mypy . → clean, 179 source files. uv run ruff check . and ruff format --check . → clean.
• Zero production changes — git diff --stat -- src/ is empty.
• Each test was mutation-verified against the exact production line it covers, then reverted: the --json payload call, total = len(previewed_groups), and the isinstance arm of the re-raise split. Every mutation produced a failure naming the behavior under test.

Disclosures

No review receipt. The lens review could not complete: the reviewer agents are provisioned with Read, Grep, Glob and no shell, while their protocol mandates gentle-ai review inspect-candidate and forbids substituting live files. Three launches returned inspection: incomplete. No PASS was authored in its place. This is a test-only change verified by the suite, lints and mutation testing above.

Pre-existing defect found while writing test 3, not fixed here. render_summary at src/openkos/cli/curate.py:1217 emits f"{stage.name}: {outcome.notice}", but all six notices already embed the stage name (curate.py:467,654,854,1089,1148,1200). Every summary line the operator sees is double-prefixed — Identity: Identity: applied 1, skipped 0. Substring assertions in the existing tests hide it. Filed separately.