Accept the daemon's canonical reply for the shell's own folder open

[coneilen]

Summary

Beta7 failed the attended Dev Box ProductionCoreProjectVisibility gate. Native
Ctrl+O registered Core exactly once in the production daemon (the CLI lists
Core C:/GraphCode-Fixtures/Core), but the shell never showed the project row.

Root cause: the shell discarded the daemon's own reply to its open request.

• Path spelling: the native picker returns C:\GraphCode-Fixtures\Core. The
  production daemon canonicalizes it to C:/GraphCode-Fixtures/Core and
  answers the openProject request with that spelling. onFrameWithEffects
  checks the incoming graph path against the pending open byte-for-byte
  (Wire.isCurrentGraphPath). So it dropped the reply and left the backslash
  pending open in place, and every later graph frame for that project was
  dropped too.
• Request ID case: the daemon echoes request IDs in uppercase (Swift
  UUID.uuidString), while the shell generates lowercase IDs. The App-level
  v2 correlation compared them byte-for-byte, so real open rejections were
  never correlated either. DaemonClient.completeRequest already compared
  them case-insensitively.

Why earlier gates passed: every fixture echoed the shell's own path spelling
and lowercase request ID, and the scrubbed registered-project case registered
through the CLI, which restores the daemon's canonical spelling. The real-daemon
round trip, the only test that speaks the production wire format, was not run
by validate.ps1 or CI.

Changes

• App.adoptCanonicalOpenPath: when a graphChanged reply is provably for
  the shell's own in-flight open, it adopts the daemon's canonical path before
  the stale-frame filter runs. On v2, "provably" means the case-insensitive
  request ID matches. On v1, which has no IDs, it means the paths are equivalent
  apart from separator and ASCII case (Wire.sameLocalProjectPath).
  Uncorrelated publications are still filtered.
• The v2 open-rejection correlation now compares request IDs
  case-insensitively.
• DaemonRoundTripTests asserts the production contract the fix relies on:
  • a backslash openProject path is answered by a graphChanged response
    with the same request ID (ignoring case);
  • the response carries the canonical forward-slash path.
• validate.ps1 now runs DaemonRoundTrip.Live.Tests.ps1 against the built
  production daemon in Windows shell integration, and a WindowsShell.Tests.ps1
  contract keeps it there.

Test plan

RED: zig test src\App.zig ... --test-filter "graph publication" with pinned Zig 0.15.2 -> 8 passed, 2 failed; canonical-path reply dropped (current graph stayed B), uppercase rejection left pending open set (expected 0, found 1)
GREEN: zig test src\App.zig ... --test-filter "graph publication" with pinned Zig 0.15.2 -> All 10 tests passed
REGRESSION: zig test src\App.zig full suite -> All 744 tests passed; zig test src\Wire.zig -> All 153 tests passed; DaemonRoundTrip.Live.Tests.ps1 with the beta7 graphcoded.exe -> REAL_DAEMON_ROUNDTRIP: PASS

Additional evidence:

• Real-daemon round trip: the new request-ID assertion first failed
  against the real beta7 daemon. It expected
  00000000-0000-4000-8000-c5cc31d533d1 and found
  00000000-0000-4000-8000-C5CC31D533D1; that is how the case defect was
  found. After the fix it printed DAEMON_OPEN_CANONICAL in all three phases.
• Scrubbed live gate: a source-built ReleaseSafe shell built with
  -Dworktrees-deferred=true, run against the beta7 daemon and CLI, gave
  SCRUBBED_SHELL_STARTUP: PASS; executed=5.
• Full local run:
  validate.ps1 -Task windows-shell -SkipTrayLive -ShellValidationRoot C:\v7
  -> unit 53/53 sections, App 744/744, scrubbed startup 5/5, and the new Real daemon wire round trip PASS. The run then failed in the local UIA gate: the attribution added in Make UIA gate popup activation deterministic and attribute wait timeouts #631 shows WinDirStat, an unrelated operator app, holding foreground at "project-row New Loop". This is recorded as a local desktop failure, not a pass.
• Exact-head CI run 37389038155 passed Windows shell integration, including the new round-trip step (3 DAEMON_OPEN_CANONICAL assertions and REAL_DAEMON_ROUNDTRIP: PASS), scrubbed startup 5/5 and the live UIA gate, along with unit shards and packaging.

Limits and honesty:

• Live native picker: I could not drive the native folder picker
  deterministically on this shared desktop. A UIA-pattern probe closed the
  dialog without returning a folder in all four attempts. No local live picker
  run is claimed. The attended Dev Box rerun on a new candidate is still
  required.
• Beta7 fresh-relaunch observation: it was a single snapshot at about 5 s
  that contained no ListItem at all, not even the Default workspace item
  every shell exposes. It is inconclusive about project rows. Locally, a fresh
  beta7 shell restores a daemon-registered Core row in about 0.5 s.
• Process note: during this investigation, one probe run accidentally
  registered a scratch fixture in the operator's real ~\.graphcode, because
  the Swift daemon ignores a USERPROFILE override. Those two files were
  removed. No such probe is part of this change.
• macOS: make test and make check were not run. The change is limited to
  the Windows Zig shell and Windows validation scripts.

Checklist

• I have read the Contributing Guidelines
• I have signed off my commits (git commit -s) per the DCO
• Tests pass locally (make test)
• Code follows the existing style (make check)
• I added the test/contract before the implementation and observed the intended RED failure