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

[vyncint]

Upgrades the PTY harness from 0.9 → 0.10.1. The break itself was small: drag takes four column-first arguments now instead of two coordinate pairs, so a transposed find result cannot be handed to it by mistake.

134 tests pass. fmt, clippy --workspace --all-targets --all-features, and the full suite are clean locally.

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

Every screen assertion in this repository reads a grid a VT emulator produced from the binary's bytes. If the binary emits a sequence that emulator does not implement, the grid is quietly wrong and every assertion becomes a claim about a plausible-looking fiction. 0.10 made it checkable, and the measured answer is one SGR 59 — underline colour, which changes no cell. Pinned exactly, so anything joining that list has to be read before the suite is trusted again.

The four blink/strikethrough entries that termlens#320 reports as unsupported despite being implemented are named in the test as the known false positives they are, rather than pinned as if real.

A coverage gap this found

crates/oxidelake-runtime/tests/oxide_tui_pty.rs is new. oxidelake-tui is tested through oxidelake-tui-demo, a binary that exists for the tests — but what people install is oxide, and oxide tui renders the same model behind clap, #[tokio::main] and a tracing_subscriber writing to stderr. tests/cli.rs drives the other subcommands with assert_cmd, which captures pipes and therefore cannot see that in a terminal stdout and stderr are one stream: a log line lands on the dashboard, and a program that never restores the terminal leaves the user's shell in the alternate screen. Only a PTY test can see either.

The rest

• crates/oxidelake-tui/tests/termlens_cli.rs drives termlens-cli at the version the lockfile names, so the tool and the library are one release.
• 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. skill-version is registered in ci-policy.py's always-run set, with tests.
• CI writes TERMLENS_ARTIFACT_DIR and renders failures into the job summary via termlens's report action (SHA-pinned).
• __pycache__ is gitignored — the Python scripts under .github/scripts have always produced it and nothing excluded it.

[vyncint]

Updated after adversarial verification. Two independent verifiers re-ran every gate and audited the diff; both returned problems, and two were real:

1. crates/oxidelake-runtime/tests/oxide_tui_pty.rs would have shipped broken. cargo package -p oxidelake-runtime --list included the test but not the snapshot it reads — that file lives in the sibling oxidelake-tui crate, which is deliberate (two independent snapshots of the same frame would drift together silently) but does not exist once this crate is unpacked on its own. Anyone running cargo test on the vendored source would have hit a missing file, and neither test was #[ignore]d. Fixed by excluding the file from the published archive: it still runs in the workspace (2 passed), and cargo package --list no longer carries it.

2. ADR-0011 and CHANGELOG.md disagreed about the starting version — the ADR said "0.6 → 0.10", the changelog "from 0.9". Both were defensible readings (the ADR's own body still recorded 0.6.1 from when the decision was taken) and a reader could not tell which was right. The ADR now says what happened: the manifest was at 0.9, and the ADR's record had gone stale.

Also gitignores __pycache__, which the Python scripts under .github/scripts have always produced and nothing excluded — it turned up as an untracked directory while staging.

Gates re-run after the fixes: fmt, clippy --workspace --all-targets --all-features, 134 tests, and check-skill-version.sh all clean.