test(session): run the load-failure and watcher-grace paths instead of grepping them

[PathGao]

What this is

Second slice of the source-shape → behaviour conversion (after #564). Test files only, no source changes.

• documentLoadFailure.test.ts — rewritten to drive a failing read through a real session.
• documentWatcherSession.test.ts — deleted; the half worth keeping moves into externalChangeReload.test.ts, which already tests this subsystem for real.

Mechanism

Both files described documentSession.svelte.ts instead of running it, and the descriptions were weaker than they looked.

Load failure. The whole assertion was:

assert.doesNotMatch(session, /tabManager\.closeTab/);
assert.match(session, /options\.onError\('Error loading file', error\);/);

The first passes for any version of the file that does not contain those characters — including one that closes the tab through closeTabs, or through the tab list, or after a rename. The second pins a string literal. Neither looks at the thing that actually matters when a read fails while another process is replacing the file: the buffer the reader could still rescue, and the recent-file entry that is their way back to it.

It now opens a tab, makes every invoke throw, calls loadMarkdown, and checks the tab is the same tab, the text is unchanged, no recent entry was deleted, and exactly one error arrived with its cause intact rather than flattened to a string.

Watcher grace. The deleted file matched three lines inside shouldReloadExternalChange. Its subject — our own writes suppress the reload — was already covered as a real behaviour test one file over (our own writes still suppress the reload, since #—). So it was duplicating a stronger test with a weaker one.

What nothing covered was the other half: that the suppression expires. That is the half with teeth. A guard that never lifts turns Live Mode off for that file for the rest of the session, and reports nothing. The new test drives it with selfWriteGraceMs: 0 — the window is already over when the event arrives, which is what a real event a second later looks like — and asserts the second consultation agrees, so the expired entry is dropped rather than re-read.

Scope

• makeSession gains an overrides argument in both files. The other eleven session tests each build their own session the same way; a shared harness for two more call sites would be a bigger change than the one being made, so none is introduced.
• The source-shape assertions at the bottom of externalChangeReload.test.ts stay. They cover wiring inside MarkdownViewer.svelte, which this runner cannot import, and the file already says so under its own heading.
• No source file changes. The two edits below were made to check the tests fail, then reverted.

Tests

Both directions were run, not reasoned about:

change to documentSession.svelte.ts	result
catch block closes the tab and deletes the recent entry (the pre-fix behaviour)	both load-failure tests red
shouldReloadExternalChange returns a bare false (suppression never expires)	expiry test red, other 10 green

Verification

npm run check   # 673 files, 0 errors
npm test        # 943 pass, 0 fail

Not verified: the tests drive documentSession with stubbed invoke, so what Rust actually returns on a vanished file is still taken on trust — as it is in the eleven existing session tests.

🤖 Generated with Claude Code