[epim]

Release 0.3.50: the version bump and the licence delta review for plan 2, waves 7 to 11.

What changes for the operator

Autofocus

• A sweep that proved nothing no longer costs the stored focus calibration; a failed sweep is retried once over a smaller span, and the focuser stays at the best point found (A sweep on a lightless sky discarded a good autofocus slope; the next night's default-span sweep found focus at HFR 2.88 and was still reported as failed #1077, Rig: focus_calibration.json was emptied on 2026-10-19, so every autofocus uses the 350-step default span and none completes #1298).
• No sweep is started right after a "no light" verdict or inside a cloud hold.
• The curve check no longer rejects a good V on roughness alone (Autofocus curve_verdict refuses a clean wide-stepped V: its roughness gate measures the sampling, not the scatter #1181).

Runs and holds

• A run that centres by plate solve is refused at the start when the rig has no plate solver, with a sentence saying what to install (#596 shape b no-light hold treats a MISSING plate solver as 'no light yet' and holds a single target for up to its whole window #721).
• A failed solve is judged on the frame it ran on, not on whatever frame was last on the hub (Hold classification reads hub.last_frame, not the frame the failed solve was run on #1182).
• Star detection, the cloud score and the no-light check measure per region, so a half-blocked frame is not averaged into a clear one (A partial obstruction skews the whole-frame statistics in detect_stars, cloud_score and the no-light check #1145).
• Only a dawn ending says "stopped at dawn"; a clock stop, a run cap or a set target says so (A run whose last window was closed by a stop time or a max_run_min cap ends saying 'stopped at dawn (windows closed)' #1239).
• The guide verdict reads the guider's own sample trace and holds for a silent guider (Guide verdict follow-ups (WP 1.2): 100-sample window, silent guider at start, require_guiding in the hold, rotation warning for PHD2/NINA #1146, The guide-RMS reject gate judges a frame on a single guider sample; a guider stall can reject a well-guided frame #1222).
• The saved-lights solve follows the solver installed now, without a restart (solve_saved_lights follows the plate solver only at config load; installing ASTAP while running changes nothing until a restart #1206).
• A warning that depends on the site now pages a fixed sentence instead of staying silent, so expect somewhat more pages (A warning flagged site_derived never pages: the flip-point cloud-hold stop is silent while its neighbours page #1165).

Flows

• Every node shows a live word: SLEW, AUTOFOCUS, GUIDE and DUSK FLATS are named, a stage left for later reads deferred, a held stage waiting, and the final status leaves nothing running (Live stage status: a resumed night publishes none (stage ids not stored), lane nodes are unnamed, a deferred stage stays 'running' #1231).
• A session created on 0.3.49 and auto-resumed after the upgrade shows no words for the new nodes until CONTINUE.
• The fixed-night angle tolerance is tighter for newly compiled mosaics that skip panels (Fixed-camera angle tolerance is slightly loose for a grid with skipped panels #1175).
• Phone: the stage-editor sheet has UNDO and REDO (Phone flow editor: the nested stage-editor sheet has no undo, and the UNDO/REDO row was checked only in jsdom at phone width #1153).
• Tablet: the flow header fits at 700 px; the backend pill is not repeated in that row (Classic FlowHeader row is estimated to overflow at the bottom of the tablet tier (700-770 px) while the save pill reads UNSAVED EDITS #770).
• Reports and the Files sheet tell two same-named mosaics apart by folder (Two same-named mosaics' bundle group rows read identically in both UIs; the row does not show the folder #1205).

Panorama scanner

• The latency estimate is right on a pan that reverses (Panorama scanner: the latency estimator reads 118.8 ms against a true 50 ms on a reverse pan #1056).
• A horizon bin is published as Measured only where a sky-to-ground step was seen; the rest is Unknown. This halves the false-open bins on the test scenes; Panorama tracer publishes false-open Measured bins on the chart yard away from any step (reads 25 where the truth is 48 to 51) #1058 stays open for the rest.

Rig backends

• ASIAIR: preview frames are deleted on every ending, and abort does not wait behind a cancelled call (ASIAIR preview frames are left on the box's SD card when an exposure ends after the save but before a successful decode (abort mid-download, failed download, failed decode) #1193, ASIAIR abort and halt still queue behind a cancelled command RPC until its libasi thread returns (bounded by the socket timeout, about 10 s per stalled read) #1194).
• The status poll does not run between a reconnect's disconnect and connect (An operator reconnect can fail when the status poll lands between its disconnect and connect #1173).

Packaging

• The server package declares its licence, Apache-2.0 (server/pyproject.toml declares no license, so the installed astrodeck distribution and the image report carry no licence metadata #1154). WCSLIB records match what ships (WCSLIB removal (#638) left stale records: RELEASE.md, frozen baseline, dated evidence, mutation evidence, GCC notice version #1159).

The licence delta review

44 shipped paths changed since their last review (27 server, 14 UI, one Rust file with comments only, the frozen-binary spec and its payload filter). Each diff was read in full and approved with its own sentence through release_evidence.py approve. None is a dependency manifest, lockfile, asset, font, image, binary or catalogue; no notice or policy file changed since 0.3.49. Nothing is withheld.

• release_evidence.py status: clean.
• Local source gate: PASS, 588 members, 0 blockers (42 before).
• Licence tool tests: 113 pass; the 16 related server test files: 208 pass.

Frozen binaries stay gate-held as before (#1073).

🤖 Generated with Claude Code

https://claude.ai/code/session_01K7CHb44VFtZLGUDLF4us2p