fix(lock): distinguish held from cannot-be-created in acquireLock (#999)

[Jay0130-a]

Summary

acquireLock returned false both when another process held the lock and when the lock could not be created at all (a parent path in the way, an unwritable directory, a full disk). Every caller read the second case as "busy": pull printed sync in progress elsewhere and exited 0 while nothing had synced, retry loops burned their whole budget, and the real cause was never named.

This splits the verdict into acquired | busy | notWritable (LockResult / acquireLockResult, with acquireLock kept as the boolean wrapper) and migrates every caller. A notWritable verdict names the lock path and reason (the closest existing path when a parent is in the way, otherwise the errno), stops retry loops immediately, and surfaces as an error on pull/push (exit code 1) instead of a silent skip.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Test Plan

• npx tsc --noEmit passes
• npm run lint — oxlint on all changed files is clean. The --type-aware pass cannot run in this checkout (it reports Failed to find tsgolint executable, i.e. the oxlint-tsgolint dev dependency is missing), not a finding.
• npx vitest run — the full suite does not complete green on this Windows machine for six pre-existing reasons below; CI on Linux is the real gate.
• Added/updated tests for the change

Related Issues

Fixes #999

Notes for Reviewers

• The notWritable shape reuses NotWritableError and existingAncestor from git-exclude.ts (the pattern fix: delivery and local-state bugs (#993) #995 established), so the message style is already familiar and tested.
• pull: a scope whose sync lock is not-writable is dropped from every clone-consuming stage exactly like a contended scope, reported with the path and reason, and fails the command (process.exitCode = 1).
• push (both self-mode and git-mode) already fails on contention; it now reports the not-writable reason instead of the generic "another process" line.
• Polling callers (votes, models/switch, models/profile, pending-learnings, dashboard-collector, jsonl-store, session-collector, secret-store, codex-team-hooks, local-agent, branch-worktree, uninstall) stop immediately on notWritable instead of waiting out their retry budget; bootstrap, migrate, source, and learnings-publish keep their existing return semantics but name the real cause.
• The EEXIST from the exclusive-create path still means "the lock exists" and stays busy; only the ensureDir-time EEXIST (a file where the lock's directory should be) is classified not-writable.
• Verification on Windows 10 / Node 24, branch @ 4c1ab4d: tsc --noEmit clean; oxlint clean; vitest run over update (64), lock-atomic (22), migrate (62), votes, model-switch (71), model-profile (30), jsonl-store, session-collector, save-session, import-lock, dashboard-collector (155), push-pending-pr (29), codex-hook-rules, self-worktree-rebind, reports-reader-lock — all pass.
• Pre-existing Windows-environment failures, byte-for-byte the same on unmodified main @ 2cbf556 in this environment: git-exclude / mcp-git-exclude (8.3 short-path vs long-path assertions), pending-learnings (POSIX path separators), secret-values / mcp-secrets / source (case or physical-target path shape), local-agent token 0o600. None are touched by this change.