test: replace fixed 50ms spins with condition waits in shortcut routing tests

[arzafran]

Second slice of #186 — but not the one the issue proposes. See below.

Why not the issue's proposal

#186 proposes extracting a pure routing decision and retiring the 87 tests. That is blocked, and worth recording rather than rediscovering.

#208 already extracted the only cleanly-pure logic (ShortcutRouting.matchStroke/match) — that's 3 table tests. The other 84 exercise routing decisions whose implementation is ~55 flat if matchConfiguredShortcut(…) { perform effect; return true } branches, with decision and effect fused per branch and no intermediate value to assert against. There is nothing pure left to table-test them against, which is the same wall that made ShortcutRouter not worth attempting in #187.

So this targets the issue's actual cost driver instead.

What this does

59 occurrences of RunLoop.main.run(until: Date(timeIntervalSinceNow: 0.05)). Each burns the full 50ms whether the work took 1ms or 49ms, and fails when it takes 51. That flakiness is precisely why ci-run-unit-tests.sh runs this class serially and absorbs a retry.

The file already diagnosed this. Three sites carry comments reading "a single fixed 0.05s spin is occasionally too short under CPU contention... causing rare flakes. Poll instead of assuming one spin is enough." This finishes what those three started.

Adds a waitUntil helper — polls the condition, returns the moment it holds, fails after 2s with the caller's file/line. 26 sites converted to wait for exactly the state the following assertion checks.

What was deliberately left alone

No assertion was added, removed, or modified. Verified: zero XCTAssert lines on either side of the diff.

27 sites stay as fixed spins, because converting them would be wrong rather than merely unfinished:

• 15 makeKeyAndOrderFront sites — a plain NSWindow can never genuinely become key in the headless test host; the file documents this. Polling isKeyWindow there would time out on every CI run. These were converted first, then reverted on finding that comment.
• 5 negative assertions ("this count should not have changed") — the spin's job is to give latent async work a window to misbehave before asserting it didn't. Polling returns instantly and defeats the test.
• 1 tearDown drain, 2 settle spins, 1 shared close helper, 1 float-accuracy assertion — none reduce to a condition.

Not touching CI yet

The serial isolation, the permanent skip, and the retry in scripts/ci-run-unit-tests.sh all stay. Removing them needs evidence from several green runs, not one. That's the follow-up once this has some history.

Test plan

• xcodebuild -scheme programa-unit build-for-testing → ** TEST BUILD SUCCEEDED **, zero error: matches, modified file confirmed in the compile log
• Zero assertion lines changed
• Fixed spins 59 → 34 (26 converted; the helper itself contains one)
• CI green — the 87 tests are their own regression check here