Stop resolution.Empty racing a store scan, and stop tests writing rows into the reports

[kyleve]

Fixes the red main (run 30402846712), plus a reporting bug found while diagnosing it. The snapshot job's only real failure was ResolutionViewSnapshotTests/resolution(), on Empty_iPhone, at max channel delta 255 over 91.7% of the image. Everything else in that run was noise the new tolerance absorbs or the already-known swiftDataInspector issue.

Merged with #150 (demo mode) — see Merging alongside #150 for the one conflict and why both sides had to survive it.

1. resolution.Empty never rendered the empty state

PreviewSupport.resolveModel(seededWithIssues: false) skipped setDataIssues entirely, so the fixture came back with hasLoaded == false — which ResolutionView deliberately cannot distinguish from "the first scan hasn't landed yet". The view therefore held AppIconLoadingView until its .task(id:) scan of the empty in-memory store returned, which found the whole year missing. So the committed reference was that scan's output — a populated list titled "Missing days / Jan 1 – Jul 14 / 195 days" — not the all-clear state the case is named for, and every capture was a race against an async store read. CI lost that race and baked the placeholder.

The settle loop can't catch this, which is why the suite had been reporting green: a placeholder is perfectly pixel-stable, so the loop proves stability and the capture proceeds. Same failure class as root.LoggedIn baking the launch splash, already ledgered in Where/TODOs.md. It stayed hidden until now because drainInFlightAnimations used to waste ~1s per capture; removing that waste in #151 took away the accidental slack. The prior main run confirms the shape — resolution() passed there in 23.0s, and failed here in 12.1s.

Fix. Both fixture modes now seed. setDataIssues([]) is what marks the model loaded and isSeeded, so the view's load(...) is a no-op and the first rendered frame is already final. The case no longer depends on the store, on now, or on how fast the machine is — there is no phase transition left to race. The two references are re-recorded to the "All clear" state, coverage the suite never had (WithIssues already pins the populated list), and consistent with the app's other empty states.

ResolveModelTests gains two guards, both deterministic and independent of the image suite: the fixture comes back loaded in both modes, and load(...) leaves a seeded fixture alone against a store whose scan does find issues — the short-circuit the empty case's determinism now rests on.

2. The reporting tests were writing rows into the reports

Found while explaining a suspicious line in ./test --review. ./test recovers SNAPSHOT_DIFF and SNAPSHOT_TIMING by grepping them out of the run logs, and counts timing lines as captured images for the progress line. Both were encoded and printed by a single function, so the tests that pin those wire formats emitted genuine lines:

• --review listed a reference that does not exist, at the top of the list, because the fixture's numbers were borrowed from the real swiftDataInspector regression (max delta 203, 7430 pixels, 0.235%). The row most worth investigating was the one that wasn't real.
• --timings counted five captures that never happened. ./test --only SnapshotKitTestingTests --timings reported "5 captures, 0.1s total, 0.024s per image" for a run that captured nothing, naming fixtures (repeated, one-phase, mixed, …) in its slowest-captures list. On --everything --timings those blended into the aggregate the module's performance decisions are read off.

The SNAPSHOT_DIFF env gate never covered this: it's checked by the pipeline before calling report, not inside it.

Fix. Each channel is split so printing belongs to the pipeline alone and the payload is separately askable — SnapshotDiffReporting.line(describing:), SnapshotCaptureTiming.line(), and SnapshotSettleReporting.line(...) return the same JSON without emitting, and all four test sites call those. The settle channel gets the same treatment for symmetry; nothing aggregates it today, but it's one scraper away from the same bug. The rule is now an invariant in AGENTS.md, since it's enforced by convention rather than by the compiler.

Verified in both directions: a unit-only run reports no timing lines and no differing captures, while the full suite still prints the complete phase breakdown — now 264 captures rather than 269 — and the diff table.

Merging alongside #150

#150 edited the adjacent lines of the same function for an unrelated reason, so neither side could be taken wholesale:

• Add demo mode, and open nothing until the user picks a world #150 replaced WherePreferences() with its new previewPreferences() helper, part of its sweep of fixtures that reached the host's real UserDefaults. This branch briefly carried a narrower version of that same fix; it is reverted here, because Add demo mode, and open nothing until the user picks a world #150's is better — it covers every fixture and removes the parameter defaults so omitting one is a compile error.
• This branch replaced if seededWithIssues { … } with setDataIssues(seededWithIssues ? sampleDataIssues() : []). Add demo mode, and open nothing until the user picks a world #150 left that line alone, so main carried the race until this merge.

The merge keeps both, and loadLeavesASeededFixtureAlone uses the makePreferences() helper #150 added to the test file.

Docs

SnapshotKitTesting's README claimed .settled "waits for pixel-stable renders so .task-driven content loads" — the over-promise that let the first bug land. It now says pixel stability gives async content time to load but cannot certify that it did. AGENTS.md gains two invariants: a settled capture is not a ready capture (seed the fixture or gate on onReadyToSnapshot), and only the pipeline may print a report channel.

Testing

Run locally on iPhone 17 / iOS 27.0, on the merged tree:

• ./test --everything --review --timings — 1365 unit tests and 30 snapshot tests / 264 images, all passing; every remaining diff is max delta 1–2 sub-visible drift and no fabricated rows appear in either report
• both reporting fixes verified from both sides: unit-only runs report nothing, real snapshot runs report fully
• swift run bumper config/test/lint . — no architecture violations
• ./swiftformat --lint and ./xcstrings --lint clean