test: termlens 0.10.1, the CLI, and the invariant the suite rests on

[vyncint]

Upgrades the harness from 0.9 → 0.10.1 and puts the two things that ship beside it — the CLI and the skill — under test. 240 tests.

The upgrade cost three lines

0.10 spread drag's two coordinate pairs into four column-first arguments, so a transposed find result can no longer be handed over by mistake. That was the entire breakage; the other 236 tests passed unchanged.

tests/emulation.rs — the invariant everything else rests on

Every other file here asserts on a grid a VT emulator produced from mossaic's bytes. If mossaic emits a sequence that emulator does not implement, the grid is quietly wrong and every assertion in this repository is a claim about a plausible-looking fiction. 0.10's Screen::unsupported made that checkable, and the answer is clean:

mode	dropped
text cells	["^[[59m"]
kitty	["^[[59m"]
sixel	["^[[59m"]

One SGR 59 — underline colour, which changes no cell. The list is pinned exactly, because anything joining it is a sequence that might change one. The image paths matter most here: they put bytes on the wire that no cell shows, so a drop there is invisible in every other assertion.

It also pins three things nothing else would notice: insert mode off (it would shift every row right and still look like a chart), no bells, and no wrapped rows (a wrapped row means the layout overflowed and the text is silently on two).

Plus round trips: a chart survives the snapshot text format byte for byte and JSON, with the palette intact — checked against a truecolour cell, since a round trip that dropped colour would still pass a text comparison.

tests/cli.rs — termlens-cli against mossaic's own screens

Saved, rendered to SVG with every truecolour the chart chose still in it (read off the live screen first, not hardcoded), and diffed. Plus inspect pointed at the real binary.

#[ignore]d: a cargo test that installs something behind a contributor's back is a surprise a published crate should not spring. CI runs it by name — the same arrangement smoke.rs already uses for the tests needing gh.

The skill, and a check so it cannot rot again

Refreshed to 0.10.1 (upstream vyncint/termlens#322 corrects the marker, since the file already described 0.10.1 behaviour). check-skill-version.sh asserts the copy and the dependency agree on major.minor. Nothing can diff it against upstream — the published crate does not ship the skill — but version agreement is the drift that actually happens. Verified it fails in both directions and tolerates a patch bump.

CI

TERMLENS_ARTIFACT_DIR on the test steps and termlens's report action on failure, so a red PTY test arrives as a picture in the job summary instead of a grid in a log. The action is SHA-pinned like everything else here.

One finding, filed not fixed

#102 — the event loop draws on every 80 ms tick whether or not anything changed, so an idle mossaic reports 244 repaints over a 20-second wait against 2 after load. Not a correctness bug (DEC 2026 is what stops it flickering), but it costs a wakeup and a write every 80 ms, and it makes this suite's wait_frame discipline load-bearing rather than belt-and-braces. Changing the event loop deserves its own consideration.