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

[vyncint]

Upgrades the PTY harness from 0.9 → 0.10.1 (with the serde feature) and adds the checks 0.10 made possible. Last of the four ecosystem repos; oxidelake, launchbound and reconverge are done.

oxmera-cli: 25 passed, 4 ignored (the CLI suite, run by name). fmt and clippy --workspace --all-targets --all-features clean.

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

Every screen assertion in this crate reads a grid a VT emulator produced from the binary's bytes. A sequence that emulator does not implement makes the grid quietly wrong and every assertion a claim about a plausible-looking fiction. The measured answer is one SGR 59 — ratatui's underline-colour reset, which changes no cell — at both geometries the suite uses (100×45 and 80×30, each a fresh spawn). insert_mode, the bell counts and mouse_modes are pinned beside it.

The rest

• tests/report.rs and tests/styles.rs assert what the dashboard says in colour alone — the loss and accuracy sparklines are pinned to their own colour per row, so a swap fails on the offending row by name rather than as one changed cell inside a whole-screen diff.
• tests/termlens_cli.rs drives termlens-cli at the version the lockfile names, so tool and library are one release. #[ignore]d, because oxmera is published and a cargo test that installs a tool behind a contributor's back is a surprise a published crate should not spring; CI runs them explicitly.
• The vendored agent skill is refreshed to 0.10.1, and check-skill-version.sh fails when that copy and the dependency disagree on major.minor — it had drifted two releases with nothing to notice.
• CI writes TERMLENS_ARTIFACT_DIR and renders failures into the job summary via termlens's report action (SHA-pinned).

Two notes

The 11 GPU tests that fail locally are pre-existing and unrelated. I verified this rather than assuming it: the same 11 (matmul_matches_cpu_including_batch_broadcast, upload_download_roundtrip, fused_adam_step_… and the rest) fail identically on a clean main worktree at 60a0802, untouched. They are device tests on a box with no GPU.

One thing needs a repository admin, not a commit: the new skill version CI job has to be added to this repo's branch-protection required-checks list, or it can fail without blocking a merge — which would make it decorative. Flagging rather than assuming it is handled.

What the adversarial review changed here

Two verifiers audited this before it became a PR. One real finding: styles.rs ended with an assert_ne! that the loss and accuracy sparklines are different colours — but the two loops directly above it had already pinned each series to a specific colour, so comparing those same cells could not fail. Removed, with the reason recorded: a swap already fails in the loops, by name and on the offending row, which is the better failure anyway.

They also caught two claims I would otherwise have put in this description — that the 80×30 measurement followed a live resize (it does not; emulation.rs spawns fresh at each geometry) and that the upgrade fixed seven breaking changes (termlens 0.10.0 lists six). Neither appears above.