release: v0.10.2 — the CLI is verified, and unsupported() stops contradicting the cell

[vyncint]

Closes #326 and #320, then cuts 0.10.2.

Why a patch now

#310 is fixed on main and has never been published. termlens <subcommand> --version exits 2 in the published 0.10.0 and 0.10.1. It was found by a user of the published binary. That alone makes this release overdue; the README restructure (#337, #338) and the skill/CI commits ride along.

#326 — the published CLI is now installed and exercised

install.yml asked "does cargo add termlens work for a stranger?" and never asked the same of the binary. A cli job now installs termlens-cli from crates.io on Linux and macOS — every run, never from a cache, because a cached binary is evidence about whatever the registry served the day the cache was filled — and asserts every exit code termlens --help documents:

assertion	expected
--version in all four positions, same string	0
inspect prints the screen, its header and its trailer	0
render --text/--svg/--html/--ansi, --svg starting <svg	0
diff, same picture	0
diff, different pictures	1
diff/render, input it cannot read; missing file; unknown subcommand	2

The assertions are a script that takes the binary, not a block inlined in the workflow, so the same contract runs against a path build. That is how it earned trust: deleting one match arm to re-create #310 made it fail on diff --version — exit 2, expected 0 — and restoring the arm made it pass.

ci.yml runs it on every pull request against target/debug/termlens. The published check is the net; this is the fast feedback, and #326's whole complaint is that the failure was found too late. gates-listed demanded the new commands in CONTRIBUTING §1, which is that gate doing its job.

#320 — unsupported() stops contradicting the cell

5, 6, 8, 9, 25, 28 and 29 are exactly what emu/shadow.rs exists to recover: vt100 drops them, they arrive at unhandled_csi, and the shadow parser puts the attribute on the cell anyway. So one screen said Style::blink was true and unsupported() said ^[[5m had been dropped, at the same instant — and assert!(screen.unsupported().is_empty()) could not pass for any application that blinks or strikes through.

Three tests, one per line of the done-when:

• every recovered parameter, on and off, names nothing — and the attribute is on the cell, which is what makes it a fix rather than silence;
• ^[[59m (underline colour, modelled by nobody) is still named;
• ^[[5;59m is named whole, because dropping a mixed sequence would hide a real gap.

The residue, measured rather than assumed. ^[[1;5;31m is still named although bold, red and blink all reach the cell — 1 and 31 are vt100's to implement, and this tracker knows the shadow's set, not the backend's. Narrowing that needs vt100's own SGR surface enumerated, which is more than a patch should claim. A test records it so it is a documented edge, not a surprise.

shadow::carrier and unhandled::SHADOW_SGR_PARAMS describe the same set in two places and could drift silently in both directions. A test holds them equal; removing 29 from one was observed to fail it.

The styled_json snapshot loses ^[[8m/^[[28m — that fixture renders conceal, so it was reporting conceal as unimplemented on a screen that shows it. The fix, visible in a recording.

Verified

cargo test --workspace --all-features 518 passed, 0 failed; every feature configuration (--no-default-features, decode, regex, serde, default) green; clippy in four configurations; MSRV 1.85; rustdoc -D warnings; gates-listed; skill-snippets; the CLI contract against a path build.

After merge

Tag v0.10.2. The new install.yml leg then runs from release.yml's post-publish step — so 0.10.2 becomes the first release whose published CLI is actually proven rather than assumed.

Closes #320
Closes #326

Signed-off-by: Vyncint Ng 115854244+vyncint@users.noreply.github.com