test(yield-campaign): stub ngspice discovery, run campaign tests in native CI leg

[loom-fleet-dispatch]

Closes #2675

• tests/test_yield_campaign.py: scoped ngspice discovery stub (plus KLT_NGSPICE_BINARY isolation) inside the two subprocess-stubbing helpers, so the six native-dependent tests no longer need a host ngspice. Production discovery untouched.
• ci.yml: native yield matrix entry now runs tests/test_yield_campaign.py; comment and missing-extension message updated. Freshness gate unchanged.

Tests: fresh native build (fingerprint e0fc0d8231c7127e verified), 154 passed, no skips; also passes with ngspice removed from PATH. ruff check/format clean. No CHANGELOG entry (test/CI only).

🤖 Generated with Claude Code

[loom-fleet-dispatch]

Judge verdict: APPROVED

Checked against #2675's revised scope:

1. The six subprocess-stubbed campaign tests no longer need a host ngspice. _stub_ngspice_discovery is called from both stubbing helpers. It uses monkeypatch, so it is scoped to those tests. It unsets KLT_NGSPICE_BINARY and makes which('ngspice') return the bare name. Production discovery is untouched.
2. The ci.yml native yield entry now lists yield, campaign and sensitivity tests. The comment and no_skip_message are updated, and the freshness gate is unchanged. The YAML parses.
3. No assertions or counts changed. The only other edit is a stale docstring about CI coverage.
4. Tests that need a real engine are unchanged.

Verification: I rebuilt klt_yield_native in a private CARGO_TARGET_DIR. The fingerprint check passed (e0fc0d8231c7127e). I ran the three suites with PATH=/usr/bin:/bin (no ngspice) and KLT_NGSPICE_BINARY unset: 154 passed, no skips. The campaign file also passes with KLT_NGSPICE_BINARY=/nonexistent. Ruff is clean. All CI checks pass and the merge state is CLEAN.

[loom-fleet-dispatch]

Verdict anchored to the current head — no marker had been recorded

This PR carries loom:pr, but no verdict-SHA marker was ever written for that verdict, so it was unverifiable: nothing could tell whether it still described the tree in front of it, and it would have survived a force-push undetected — the exact pre-#5686 hazard.

This comment records the head SHA as of now, c65526d8da73624bdeddc047a4b0f625a229089a. It is not a review and implies no judgment about this tree: the loom:pr label is unchanged. From here on the verdict is invalidatable — if the head moves off c65526d8da73624bdeddc047a4b0f625a229089a, the stale-verdict pass clears loom:pr and returns the PR to loom:review-requested.

Anchoring bounds future exposure; it cannot reconstruct which tree was actually reviewed. If the head already moved before this comment, treat the verdict with corresponding suspicion.

---

Automated by loom-daemon claim reconciliation (#6319)