fix(client-runtime): reopened files show host edits instead of the first read#29409

Problem

On mobile, reopening a workspace file showed the contents from the first time it was opened, even after the file changed on the host. Pulling to refresh the folder, navigating away and back, and switching threads did not help; only closing the app did. Reproduced on a Pixel 9 (Android 16) against a dev server on 701e13cebb: open proof.md (Revision A), go back, write Revision B on the host, reopen after ~38 s. The viewer still shows A, and the host sees no projects.readFile request for the reopen.

Environment queries (createEnvironmentQueryAtomFamily) asked for Atom.swr({ revalidateOnMount: true }) under a 5-minute idle TTL. Atom.swr only checks staleness when its node is built, and the TTL keeps that node alive across an unmount, so a remount inside the TTL never revalidated. Each reopen also restarted the TTL.

Change

• packages/client-runtime/src/state/runtime.ts: replace Atom.swr with a small outermost revalidateOnMount layer that has no idle TTL, over the TTL'd chain that holds the cache. Its node is rebuilt on every mount; when the cached result is older than the query's staleTimeMs it shows the cached value and refetches once, deferred out of the render like Atom.swr. A fresh result is not refetched, and nothing runs per render.
• Atom.family holds members weakly, so each family keeps a WeakMap from the TTL'd cache atom to the returned atom. The member lives exactly as long as its cache; once the TTL drops the cache node, both can be collected.
• apps/web/src/state/revisionedQuery.ts: a query counts as cached when its own node or its initialValueTarget's node exists, since the outer node now goes with its last reader while the cache stays.
• Families with a refresh interval keep their source subscribed while cached (observeWhileCached), as Atom.swr did. Without that, an idle interval refresh only marked the query stale, so a cached signed asset URL could expire before the next reader mounted.
• packages/client-runtime/src/state/assets.ts: asset URL queries use their 30-minute refresh interval as the stale time. Each mint signs a new URL, so a remounted reader reuses its URL until it is due for renewal (tokens last 60 minutes) instead of re-minting after 5 minutes.

Scope and approval

Fixes pingdotgg/t3code#29345

Verification

Focused tests:

• packages/client-runtime/src/state/runtime.test.ts
  • environment query lifecycle > refetches when mounted again after its stale time, while still cached: fails on main (expected 'before edit' to be 'after edit'), passes here. Includes a remount inside the stale time that must not refetch.
  • environment query lifecycle > keeps its cache when collected while unmounted, inside its stale time: forces a GC between unmount and remount; fails without the WeakMap (a second, cold fetch), passes here.
  • environment query lifecycle > keeps refreshing on its interval while cached with no reader: an idle interval refresh must still fetch; without observeWhileCached it never answers.
• Asset URL lifecycle, checked with a scratch witness against the production asset family: mint at 0, unmount at 20 min, remount at 70 min returns the 30-minute renewal (valid until 90 min), as on main.
• apps/web/src/state/revisionedQuery.test.ts: new case for a cache retained under its initialValueTarget; fails with the old node check.
• Also passing: all of packages/client-runtime/src/state plus revisionedQuery.test.ts and projectFilesQueryState.test.ts(x) (68 files, 790 tests), revisionedQuery.browser.test.tsx, and vp run check:changed after merging main.

Android recordings, Pixel 9, Android 16 / API 36, portrait 1080 × 2424, DPR 2.625, font scale 0.94, dark theme. Fresh dev servers (T3CODE_DEV_SEED=0) per side, seeded with the same synthetic project through the server RPC. Each lane: open proof.md (A) → Back → host shell write of B → reopen the same way ~38 s after the last read → settle. No pull-to-refresh, restart or reload inside a lane.

Settled Before/After stills differ at zero threshold only in the A/B glyph, the status-bar clock and the signal icon (M1: 1,174 changed pixels of 2,617,920). The After captures were taken at 106e5d840a; later commits change only the asset-URL stale time and interval-family refreshes, neither of which is on the file viewer's readFile path (no refresh interval). An independent rebuild of M1 at the head, from the same seed and driver, matched the delivered After still pixel for pixel below the status bar. The recordings, stills, diff images, seed and driver scripts are not attached to this PR.

Not checked: iOS (same shared code path), web file panel remounts, Hermes' WeakRef/FinalizationRegistry behaviour (without them Atom.family holds members strongly and the fix still applies). A file viewer that stays open does not refresh by itself; pull to refresh still covers that.

Model and harness: Claude Opus 5.5 (lead, tests, implementation, review) and GPT-6.1 Sol (review, proof audit, capture), in Claude Code and Codex, orchestrated through T3 Code.

🤖 Generated with Claude Code