fix(cli): let render_summary own the stage prefix exactly once (#504)

[jasonssdev]

Closes #504

The defect

render_summary (src/openkos/cli/curate.py:1213) prefixes every line with "{stage}: ". Sixteen notice sites named their own stage as well, so those lines rendered doubled:

Identity: Identity: applied 1, skipped 0.
Structure: Structure: skipped -- Ollama unavailable (see above).

The probe-derived notices — probe.unavailable (curate.py:294) and probe.empty_message (curate.py:325,513,716,912) — never named a stage, so they rendered correctly. A single summary block therefore mixed two formats depending on which branch produced the outcome:

Identity: Identity: applied 1, skipped 0.      <- doubled
Structure: No untyped edges found.             <- correct

The issue as originally filed claimed uniform doubling and listed seven sites. Both were wrong and are corrected in this comment: the defect is inconsistency, and there are sixteen sites. The original repro was misleading because it constructed notices with the stage name instead of using real ones — it proved render_summary prepends, not that real notices double.

The fix

Strip the stage name from all sixteen sites; leave render_summary as the sole owner of the prefix. This adopts the convention the probe-derived paths were already using rather than inventing one. Each notice's post-prefix text is preserved byte for byte, so output changes only by losing the duplicate.

Blast radius verified as contained:

• StageOutcome.notice is consumed at exactly one place (curate.py:1213); render_summary has exactly one caller (src/openkos/cli/main.py:11326).
• ctx.ollama_unavailable_notice stores one of these strings but is only ever tested is not None (curate.py:1127), never printed on its own.
• probe.notice (curate.py:1117) is a different object echoed raw to stderr — untouched.
• Three stage.name uses correctly remain: the stderr refusal (curate.py:241), stage_notice observability (curate.py:1100), and the prefix owner itself (curate.py:1213).

Both docstrings now state the invariant so the next notice does not reintroduce it.

Why it survived this long

All 26 summary assertions in tests/unit/cli/test_curate.py were substring checks, and "Identity: applied 1" is a substring of "Identity: Identity: applied 1". 22 are now full-line assertions through a new _lines() helper; the other 4 are not summary-line assertions (one absence check, two fixture values, one equality against a production-built outcome).

That conversion — not the prefix removal — is the part that keeps this fixed.

Verification

• TDD: the failing test came first. RED showed 'Identity: Identity: applied 1, skipped 0.' sitting beside the correct 'Structure: No untyped edges found.'
• uv run pytest -q → 4037 passed, 1 skipped (was 4035; test_curate.py 98 → 100).
• uv run mypy . clean on 179 files; ruff check and ruff format --check clean.
• Mutation-verified: reintroducing Identity: at curate.py:469 alone fails six tests, then reverted with the inverse edit.

Disclosure

No review receipt. Two independent reasons, both stated plainly:

1. The lens reviewers cannot complete in this runtime — they are provisioned Read, Grep, Glob with no shell while their protocol mandates gentle-ai review inspect-candidate and forbids substituting live files. Three launches on the previous candidate returned inspection: incomplete.
2. I committed before opening the review, so the provider froze an empty candidate (0 files, 0 lines), classified it low risk with zero lenses, and would have issued a receipt that reviewed nothing. That lineage (review-505bbb08be465499) was deliberately left unfinalized rather than presented as coverage.

No PASS was authored in either case. The change stands on the suite, the lints, and the mutation evidence above.