fix(core): CAS write gateway + per-path advisory locks for TOCTOU safety

[bigknoxy]

Problem

Closes #21

Every tier (AST, hash, diff) does read→transform→write with no re-validation between the read and the write. Concurrent edits silently lose work (last-write-wins). The hash tier's stale-anchor check defaults to auto-recovery instead of aborting.

Approach

1. CAS (Compare-And-Swap) Write Gateway

• routeRead() returns file content + hash for CAS gating
• casWrite() re-reads file hash immediately before write, aborts with STALE_ANCHOR if stale, returns fresh hash for retry
• Applies to all three tiers (AST, hash, diff), not just hash

2. Per-Path Advisory Lock (src/core/lock.ts)

• Atomic file creation (.hashpilot/locks/<sha256>.lock)
• PID-liveness stale lock breaking
• Sorted lock acquisition order prevents deadlock
• LOCK_TIMEOUT error code with bounded wait

3. Hash Tier noRecovery

• replaceHash() called with noRecovery: true when oldHash is set
• Dry-run bypasses stale check for simulation

4. editMany Conflict Distinction

• BatchSummary.conflicts counts stale-anchor failures separately from logic failures

Tests (23 new)

Category	Tests
routeRead	3 (content, hash, nonexistent)
CAS: AST tier	3 (stale abort, correct hash, backward compat)
CAS: diff tier	2 (stale abort, correct hash)
CAS: hash tier	2 (no auto-recover, correct hash)
Dry-run bypass	2 (AST, hash)
No regression	3 (AST, hash, diff)
LOCK_TIMEOUT	2
Advisory lock	3 (create/release, timeout, stale break)
Deadlock prevention	2 (sorted order, property check)
editMany conflicts	1

Sabotage run verified: reverting CAS changes → tests fail, confirming they catch the absence of the fix.

Risk

• Breaking change: Hash tier with oldHash now fails instead of auto-recovering. Callers relying on auto-recovery need to handle STALE_ANCHOR and retry.
• Backward compat: Calling routeEdit without oldHash still works (CAS is opt-in via the hash parameter).

Exclusions

• Per-path locks are process-local only (no networked locking for remote FS)
• Advisory locks use polling (100ms interval) rather than inotify/fsevents
• No changes to the CLI interface