fix(cli): bridge two vocabularies, move a late warning, stop a contradictory PASS (#389)

[jasonssdev]

Closes #389 — the last of its bullets. One needed no change; the rest are here.

Already fixed, verified rather than assumed

doctor's Ollama remediation already distinguishes "not installed" from "not running": it checks PATH and, when the binary is absent, points at the download page and mentions the macOS app instead of printing a serve command that does not exist on that machine. Traced to the commit that introduced it rather than taken from the issue text.

The three real ones

Vocabulary. reconcile reported "recorded 'X' as superseding 'Y'" while list shows deprecated in its STATUS column — two words for the action just performed and its effect, with nothing connecting them. The line now names the status the superseded concept will carry.

Placement. init's embedding stickiness note arrived after the choice it informs and below "Next: run openkos ingest" — past both moments it exists for. A note now precedes the picker; the concrete one naming the resolved tag sits above the call to action.

A green line under a red one. doctor printed [PASS] Backend host locality directly beneath [FAIL] Ollama reachable. It reports configuration, not liveness, which is correct — but the reader has no way to know the two answer different questions. It is now [SKIP] when Ollama is unreachable, following the same rule the model and embedding checks already use, with the configured host still in the detail so the fact survives.

What the reviews caught — three rounds, all of it mine

Round 1 fixed the placement and immediately created a new problem: an interactive run now printed the stickiness explanation twice, once before the picker and once after. Moving a warning earlier is worthless if the reader meets the same sentence again five lines down. A flag now selects: if the picker note carried the explanation, the later line only confirms which tag it applies to.

Round 1 also caught that a test added in that same commit was vacuous. It claimed to prove the note precedes a rendered embedding picker, but patching the installed-model probe to return an empty list makes that picker take its documented silent fallback — printing nothing at all. It passed for the same weak reason as its non-interactive sibling. It now reports an allowlisted model as installed and asserts the picker text actually appeared.

Rounds 2 and 3 were about prose telling lies: the doctor docstring still said the locality check has no skip branch, which round 1 had just made false; then my consolidation added a sentence saying the reasoning "is not repeated here" while leaving the repeat directly below it. Two lenses caught that one independently.

Verification

• uv run pytest -q → 4032 passed, 1 skipped. uv run mypy . → Success, 179 files. ruff check / ruff format clean and convergent.
• Four of the five new tests were driven RED→GREEN. The fifth was mutation-verified: forcing the locality check to skip unconditionally failed it and three pre-existing healthy-run tests.
• The one-explanation-per-run claim has its own test, mutation-verified by making both branches print.
• The corrected picker test now asserts the picker text rendered — otherwise it proves nothing about placement relative to it.

Native review

Five lineages, all approved with receipts, all zero blockers. The last one returned zero findings across all four lenses.

pre-pr denies this PR, same structural reason as #498 through #501: it measures the whole branch against origin/main while each receipt binds its own increment. receipt-binding / candidate-or-paths-mismatch, action explicit-maintainer-action. Every changed line was reviewed and approved.