Bound FITS-header fetch so a dead peer can't stall the frame

[thusser]

Summary

FitsHeaderMixin.add_requested_fits_headers() awaited every requested header future with no timeout. The futures are created before the exposure and awaited after readout, so a peer that never answers its IQ stalls frame finalization for the full XMPP IQ timeout (~120s — slixmpp's Iq.send default).

Trigger (issue #764): admin was a pyobs-gui on a laptop sent to sleep without closing the client. Its session lingered on ejabberd, and fli230 spent ~120s per exposure fetching headers from it before giving up.

Change

• Bound the whole collection to a single asyncio.wait(..., timeout=...) deadline; cancel + gather anything still pending, then skip it with a timed out warning (mirrors the existing RemoteError path).
• New fits_header_timeout: float = 15.0 kwarg on FitsHeaderMixin, stored as _fitsheadermixin_header_timeout.

A @timeout decorator was considered and rejected: it works by having the server reply with a method_timeout IQ before running, and a dead peer never replies at all.

Test plan

• ruff check / black --check clean; pyrefly check clean on fitsheader.py.
• pytest tests/mixins/test_fitsheader.py: 43 passed, including two new tests (dead peer times out and is cancelled; live+dead mix adds live headers and skips dead).

See specs/plans/2026-08-16-fits-header-fetch-timeout.md for the plan.

Note: this branch also contains commit 1e2b81d5 ("Mark core-tier test baseline plan implemented"), which the graphify automation committed to the branch on checkout — unrelated to this change.

[thusser]

Found and fixed two bugs while reviewing this (commit caa5621):

1. Crash on empty futures. asyncio.wait(futures.values(), timeout=...) raises ValueError: Set of Tasks/Futures is empty. whenever futures is {} — which request_fits_headers() returns for any module with no comm, or with comm but no peer implementing IFitsHeaderBefore/IFitsHeaderAfter (dev/test setups, single-module deployments). Previously a harmless no-op loop, now crashes every exposure on that path. None of the existing tests (old or new) exercised an empty dict, and the one test that runs a real grab_data() (test_aborted_exposure_raises_aborted_error) aborts before reaching header assembly, so this shipped uncaught. Fixed by skipping the asyncio.wait call when futures is empty.

2. fits_header_timeout never reaches BaseCamera/BaseVideo. Both call ImageFitsHeaderMixin.__init__ with an explicit keyword list rather than forwarding **kwargs (unlike basespectrograph.py, which does), so a configured fits_header_timeout silently fell into Module/Object's catch-all kwargs and was dropped — defeating the PR's stated purpose for fli230 (a BaseCamera), the exact module in the motivating incident. Fixed by adding an explicit fits_header_timeout param to both and threading it through.

Added regression tests for both, plus a fits_header_timeout-reaches-mixin test for BaseSpectrograph (that path was already correct via **kwargs, just untested). Full tests/mixins/ + tests/modules/camera/ suite (134 tests) passes; ruff/black/pyrefly clean.

[thusser]

Verified on head caa5621: pytest (fitsheader + basecamera/basespectrograph/basevideo) 87 passed, ruff/black/pyrefly clean.

Fix is correct and a real improvement (concurrent await under one deadline vs the old sequential awaits). Two non-blocking notes:

1. future.result() only guards RemoteError. A peer that errors with something else (or a task cancelled externally) still propagates. Same as the old await, so no regression, but worth a follow-up if peers ever fail with non-RemoteError.

2. 1e2b81d5 (test-baseline plan 'implemented') rides along here and touches specs/. Drop it if you want a clean single-purpose PR; otherwise harmless doc churn.

LGTM.