fix(widgets): hold builds while a dev folder is made again, and ignore folder-only change events

[mrgoonie]

Fixes #638. Follows #614 (#611).

What changed

1. No failed builds while a dev folder is made again (packages/core/src/widget-dev-engine.ts)

• Every build now goes through attempt(). A build that would start while the watched folder is missing within the 2 s grace is held, not run.
• A build that fails while the folder went, or came back under a new id, during the build is also held, and its failure is not reported. A counter of absences and re-arms catches a return that happened mid-build.
• When the folder is back, the existing re-arm (or, for the same folder returning, a new scheduled build) runs one build. That build is reported once. A held rebuild() (and the runtime's POST .../rebuild) settles with it, after onBuild has been called, so the runtime view already shows it. A held rebuild keeps its rebuild trigger.
• If the folder does not come back within the grace, nothing changes: the session stops as folder-gone. A waiting rebuild settles with the held failure, or with FILES_UNREADABLE "the folder was missing".
• An unwatched engine (watch: false) never holds, so the standalone dev host behaves as before.

2. No redundant build after a re-arm (Windows)

• Root cause, from a probe and a trace of the engine on main: the first directory listing of a freshly made subfolder makes Windows fire a change event that names the folder (change widgets, change widgets\main). This is the NTFS last-access update. Reading the files fires nothing, and a second listing fires nothing. The re-arm build lists the new folders, so it caused the next build, an unchanged one.
• Fix: the watcher ignores a change event whose name is an existing folder (checked with lstat). Adding, removing or saving a file is still reported under the file's own name, so real saves still build. The tests show this, including a save that adds new nested folders, and a save straight after a re-arm.

The runtime source is not changed. Only the runtime test that encoded the old "rebuild during grace fails" behaviour was updated.

Docs: docs/open-interfaces.md and docs/open-interfaces.vi.md describe the held builds and folder-only change events. The official docs (clarkcant-web) do not describe the grace or build results at this level (docs/cli.html only lists stop reasons), so they are not stale.

Tests

New or changed in packages/core/test/widget-dev-engine.spec.ts:

• does not call a watched folder gone while it may still come back...: a rebuild during the grace waits, lastBuild is unchanged, nothing is reported, and it settles as failed only once the folder is gone.
• holds a build that fails because the folder went while it ran...: the folder is deleted exactly as the build reads the manifest (a deterministic readFileSync hook). The result is no failure, and exactly ["rebuild:generation"] once the folder is back.
• builds a folder another process makes again exactly once...: a child process does rm -rf and then copies 500 ms later, with 4 rebuilds during the gap and a cacheRoot. All 4 settle as the same generation, and built is exactly ["rebuild:generation"].
• builds a folder another process made again once, though the build's own reading makes the platform report changes: after a child-process recreate, exactly ["change:generation"], and a save right after it still builds.
• builds a save that adds a folder, once, and a save right after it.

apps/runtime/test/widget-dev-sessions.spec.ts: a rebuild during a child-process recreate now answers live, with lastBuild ok at generation 2 and generation 2 active.

Evidence (Windows 11, Node 24.11)

• Against main's engine:
  • The 3 grace tests fail. For example, the 4 rebuilds come back as failed ×4.
  • The runtime test fails (ok: false, generation 1).
  • The redundant-build test fails in 3 of 4 instrumented runs with ["change:generation","change:unchanged"]. The trace shows the folder change events arriving during the re-arm build.
• On the branch:
  • Engine, runtime and widget-cli specs pass in 3 of 3 runs (263/263).
  • The redundant-build test passed 6 of 6 times.
• pnpm typecheck: exit 0. eslint on the changed files: exit 0. pnpm invariants: 15/15 pass.
• pnpm verify: exit 0, with 575 files passed and 1 skipped, and 7858 tests passed.

No open PR overlaps these files. #628, which is in progress elsewhere, touches the runtime chosen-folder code. This PR does not change runtime source.

[mrgoonie]

Review attestation: ready to merge at 28767ff1ff5876bfd1fd2c6d903be60a8426c1f1, reviewed by agent:code-reviewer.

A push to this PR makes this attestation stale; the new head needs its own review.