feat(terminal): record OSC 8 hyperlinks on the cells they cover

[Ayman Bagabas (aymanbagabas)]

Adds OSC 8 hyperlinks to the cell vocabulary, so a test can assert where a link points rather than only what it says.

Why

Nothing read a cell's hyperlink. EmuCell had no field for one, so every backend parsed OSC 8 and then dropped it at the seam. A CLI that prints View logs as a link could be asserted on for its text and never for its URL, which is the part that breaks.

The model

pub struct Hyperlink {
    pub id: Option<CompactString>,
    pub uri: CompactString,
}
// on EmuCell:
pub hyperlink: Option<Arc<Hyperlink>>,

Behind an Arc because a link covers a run of cells rather than one. Inline, a 40-character URL would be copied into every cell of the run and would more than double EmuCell; alacritty and rio already hand out reference-counted links, so this is also what they cost to clone.

Surfaced as link and link_id on the cell model, and as a link key in snapshots. The snapshot key records the URI rather than a boolean, on the same reasoning as the existing underline key recording a style rather than a boolean: a snapshot that only said "this is a link" would pass when the link silently started pointing somewhere else.

Only id and the URI

OSC 8 allows arbitrary key=value params before the URI. All four backends keep id and discard the rest before a cell can be read back:

backend	accessor	keeps
alacritty	Cell::hyperlink()	id, uri
rio	Extras::hyperlink	id, uri
ghostty	GridRef::hyperlink_uri()	uri only
xterm.js	oscLinkService.getLinkData(urlId)	id, uri

So there is nothing left to report, and reporting the full param list would mean parsing OSC 8 a second time ourselves in parallel with the emulator.

Synthesized ids are dropped. alacritty and rio each invent an id for a link that arrived without one, appending _alacritty or _rio to a process-wide counter. That is a rendering aid rather than something the child sent, it varies with the order links were parsed in, and no other backend produces it, so letting it through would put a backend's private counter into snapshots. A link with no id= reports None.

ghostty reports no id at all: its FFI exposes ghostty_grid_ref_hyperlink_uri and has no accessor for the parameter, so it never crosses the boundary. Declared as a hyperlink_has_no_id divergence rather than worked around.

Two xterm.js bugs, both found by the new tests

A link ran to the end of the row. getCell reuses a single CellData instance and overwrites extended only for a cell that actually has extended attributes, so an unlinked cell read straight after a linked one still carried the previous cell's urlId. Now gated on hasExtendedAttrs().

A link forced a dashed underline. ExtendedAttrs.underlineStyle returns 5 whenever a urlId is set:

get underlineStyle(){return this._urlId?5:(469762048&this._ext)>>26}

That is how xterm.js chooses to draw links, not something the child sent, and the other three backends report what SGR said. Only reachable for a cell that is both underlined and linked, since a link alone never sets the underline flag. The raw bits still hold the real style, so those are read instead.

Both were latent before this PR: xterm.js has been reporting dashed underlines on underlined links since the backend landed, with no test that could see it.

ghostty's viewport path

ghostty's render iterator exposes a cell's style and graphemes but not its hyperlink, so the URI has to come from a grid reference, which costs an FFI call per cell. Row::has_hyperlink narrows that to the rows that have one; it is documented as allowing false positives but not false negatives, so a row it skips genuinely has no link. A session with no links pays one flag check per row.

Configuration

Profile::hyperlinks, default on, plumbed through tui-test.toml, JS (hyperlinks) and Python (hyperlinks). Off makes a session behave like a terminal without OSC 8 support, which is what a program's fallback text renders on. No backend offers a native switch, so each is gated where the cell is read.

Tests

Five conformance cases, so all four backends are held to the same contract:

• conformance_hyperlinks_cover_the_cells_they_open_over
• conformance_hyperlink_ids_are_only_what_the_child_sent
• conformance_adjacent_hyperlinks_stay_separate
• conformance_a_hyperlink_does_not_change_the_underline
• conformance_hyperlinks_can_be_turned_off

cargo test --workspace --features tui-test-rs/ghostty,tui-test-rs/rio,tui-test-rs/xtermjs gives 435 passed, 0 failed.

Independent of #180; both touch Profile and the bindings, so they will need a trivial merge if both land.