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

[vyncint]

Upgrades the PTY harness from 0.9 → 0.10.1. drag taking four column-first arguments instead of two coordinate pairs was the whole breakage.

74 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 crate reads a grid a VT emulator produced from the binary's bytes — five golden files, two border scans and a styled banner. A sequence that emulator does not implement makes the grid quietly wrong and every one of those assertions a claim about a plausible-looking fiction. The measured answer is one SGR 59: underline colour, which changes no cell. Pinned exactly.

The rest

• tests/cli.rs drives termlens-cli against this crate's own saved screens — rendered, diffed with its 0/1/2 exit codes, and inspect pointed at the real binary — resolved at the version the lockfile names so tool and library are one release.
• tests/tui.rs gains coverage using the 0.10 surface, plus a styled golden and a run-metal fixture.
• 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).
• CHANGELOG.md updated under [Unreleased].

[vyncint]

Updated after adversarial verification. Two independent verifiers re-ran every gate and audited the diff. They found four things worth fixing, three of which were assertions or comments that could not do what they claimed:

1. A round trip that could not fail. tui.rs read the styled golden back with Screen::parse and asserted frame.diff(&parsed).is_empty(), under a comment saying it proved the file "describes this screen, cell for cell, rather than merely being equal to a string this run produced". But assert_styled_golden on the line above had already asserted the file equals that string — so the diff could only fail if termlens' own round trip were broken, which is a claim about the harness, not about launchbound. Removed; emulation.rs::a_frame_survives_the_snapshot_format_and_json makes that claim properly, against a live screen.

2. An assertion guaranteed by arithmetic. emulation.rs asserted a 58-character sentence was absent from any single row of a 60-column bordered frame. It could not have fit whatever the renderer did — that is arithmetic dressed as a test. Removed, with the reason recorded; the claim worth making is the one above it (no row is terminal-wrapped, so the breaking was the renderer's doing).

3. A coverage claim that overstated the loop. The doc said the invariant ran over "both fixtures, all four views and both widths", but run-metal only ran at 80x24. Fixed by adding ("run-metal", (60, 30)) rather than trimming the sentence — the claim is now true and the coverage is real.

4. A second drift hole this PR opened. cli-version: "0.10.1" is a literal in three workflow files, under a comment claiming it was "the version the lockfile names" — nothing compared them. check-skill-version.sh now checks the workflow pins against the manifest too, and I verified it fails when one goes stale.

Gates re-run after the fixes: fmt, clippy, 74 tests, and the extended check-skill-version.sh all clean.