fix: four board bugs — cockpit config, hardcoded runtime, Spotlight noindex, registry test isolation

[unfoundbox]

Summary

Four small, already-diagnosed bugs from the project board. Each is fixed and verified independently (see NOT CONFIRMED below for the one caveat).

1. spacepilot/api/routes/gpu.py — /api/cockpit/status read cfg.get("region", ...), but the config (spacepilot/cli.py's DEFAULT_CONFIG) only ever has aws_region. instance_type already matched. Fixed the key name (and the stale g6e.xlarge fallback, which should be g6e.2xlarge per the repo's actual quota). Verified by calling get_cockpit_status() directly with a fake config: it now reports the real configured region instead of always falling back to the default.

2. spacepilot/api/routes/inference.py — _extensions() hardcoded "runtime": "mlx-lm" in every /v1/chat/completions and /v1/embeddings success response. The real routed runtime is already computed once per request in _resolve() via routes.route_for() (the same lookup /v1/models reports as x_spacepilot.runtime) — it was just being thrown away after the "is there a route at all" check. Now _resolve() returns it and every call site threads it through instead of hardcoding. Verified: full test suite green (including the existing assertion that today's single wired runtime still reports as "mlx-lm"), plus a direct unit check that _extensions() reports whatever runtime it's given rather than a fixed string.

3. native/SpaceBar/tools/make_app.sh — the suggested fix (touch .build/.metadata_never_index) turned out to be based on a mechanism that (a) only ever applied at a volume root, not an arbitrary subfolder, and (b) is confirmed broken on current macOS by multiple independent sources. The mechanism that still works, documented since macOS 10.6 and used by Xcode's own ModuleCache.noindex: a directory whose name ends in .noindex is skipped by Spotlight entirely. The assembled debug SpaceBar.app now lives at .build/app.noindex/SpaceBar.app. Updated the two other places that referenced the old path (SelfTest.swift, docs/LOCAL-TEST.md).

4. Registry export test fragility — test_exported_json_matches_the_registry regenerates web/registry.json from the real shipped corpus merged with whatever a test session wrote to the default (env-redirected) measurement/system store. A test that writes there under a real/shipped system id (e.g. apple-m1-max-32gb) would silently poison that comparison, possibly on a run far away from the one that caused it. tests/conftest.py now has an autouse fixture that fails the offending test immediately if this happens, plus a fake_system/FAKE_SYSTEM_ID fixture tests can reach for. Fixed the one real-looking hardcoded id already in the suite (test_inference_surface.py, previously harmless only because it happened to also pass an explicit root=tmp_path). Added a regression test that deliberately records under a shipped id and confirms the guard catches it, then confirms the store is clean again after cleanup.

Verification

• Full suite: 806 passed, 2 skipped, 11 deselected (packaging), 0 failed via local-ml-py311 interpreter (python -m pytest tests/ -q).
• Items 1 and 2 verified with direct, real function calls (not just existing tests) — see commit body / PR description above.
• Item 3 (make_app.sh): bash -n syntax-checked, and the nested-directory logic verified by reproducing it standalone. Did not run the actual swift build — Swift toolchain compute is exactly the kind of load the fan rule exists to keep off this laptop, and it isn't necessary to prove the Spotlight mechanism is correct. Did not do a live Spotlight (mdfind) round-trip either — that's timing-dependent and unreliable to assert against in this environment.

NOT CONFIRMED

• make_app.sh: the full build-and-sign flow was not executed end-to-end (no swift build, no codesign, no live Spotlight query against a real built app). The path/variable logic was verified in isolation; the .noindex mechanism itself is corroborated by several independent current sources but not by a live mdfind test on this exact machine.

Test plan

• python -m pytest tests/ -q — 806 passed, 2 skipped, 0 failed
• Direct verification of the cockpit status region/instance_type fix
• Direct verification that _extensions() no longer hardcodes the runtime
• New regression test for the registry-export pollution guard passes
• Manual make_app.sh build + Spotlight check (not run — see NOT CONFIRMED)

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
spacepilot	Error		Sep 24, 2026 4:51pm UTC
spacepilot.dev	Ready	Preview	Sep 24, 2026 4:51pm UTC