Explain a stale latest pointer instead of leaking a raw ENOENT

[jantimon]

Defect

latest resolves through a cwd-keyed pointer file ($XDG_STATE_HOME/wpd/pointers/<hash>.json), so the recording it names can be gone by the time a verb reads it: a cleaned temp dir, a trashed artifact. Every consumer verb then printed the filesystem's own error, which names an internal path and offers no next step:

$ wpd query spans latest
ENOENT: no such file or directory, open '/private/tmp/claude-501/.../repro/rec.json'

Same for assert latest, query blame latest, cpu-diff latest latest, and a run-group whose member recording was deleted.

Fix

The check lives in the resolution layer, so every verb gets it:

• resolve.ts gains requireArtifact(absPath, subject, fix). It rewrites only the not-there family (ENOENT/ENOTDIR) into a message naming which artifact is gone, its path through displayPath (the same way every report shows one), and the recovery. A permission or IO failure keeps its own error, and so does every path the user typed, so a corrupt or wrong-schema file still gets the schema-epoch message.
• Each artifact names itself: the recording, the run-group manifest, the CPU model / raw profile, the allocation model / raw profile. query cpu latest blames the .cpu.json, not a recording sitting right there.
• Run-group members go through one checked path helper, requireMemberRecording, used by loadMemberRecording, resolveVerbTarget, the query spans/query span stitches, both assert member reads, and the per-member diff pair. memberRecordingPath stays as the pure path and is no longer exported.
• assertMemberMode also shows its path through displayPath.

After:

$ wpd query spans latest
The last recording for this directory no longer exists at rec.json. Run `record` again, or pass an explicit file path.

$ wpd query spans latest          # run-group with a deleted member
The recording for run-group member 'node-cpu' no longer exists at g1.json. Re-record the member, or pass an explicit recording path.

Exit status is unchanged (non-zero).

Tests

test/unit/stale-pointer.test.mjs (browser-free, against compiled dist/, XDG_STATE_HOME pointed at a temp dir):

• the exact stale-recording message from both resolveTarget and resolveConsumption
• the CPU model and the allocation model naming their own paths
• a stale group-manifest pointer naming the manifest
• ENOTDIR (a path under a file) worded as the same not-there case
• an explicit missing path resolving unchecked, so the reader's own error survives
• a deleted run-group member, through loadMemberRecording and resolveVerbTarget

Two existing tests pointed latest at files that were never created; they now create them, rather than weakening the new check.

Gates: build, 761 unit tests, lint, format:check, knip all pass. Verified on the real CLI for both the plain and the run-group case.