fix(pre-read): skip repeated-read warning when file was modified since last read

[AdarshJ173]

Summary

Fixes #41 — pre-read hook blocks re-reads even after the file was modified during the session.

The pre-read hook currently warns on every re-read of a file seen this session, even when Claude just wrote to that file or the user edited it externally. This inflates repeated_reads_warned and estimated_savings_vs_bare_cli stats with false positives, and — more practically — suppresses the anatomy hint and re-read token estimate that Claude needs after a file changes.

---

What changed

src/hooks/pre-read.ts

• Problem: session.files_read[normalizedFile] stored only count, tokens, and first_read — no mtime snapshot, so there was no way to detect a change between reads.
• Change 1: When recording the initial read entry, capture read_mtime: fs.statSync(normalizedFile).mtimeMs (wrapped in try/catch; if statSync fails, read_mtime is left undefined and the old behaviour is preserved).
• Change 2: On re-read, if prev.read_mtime !== undefined, compare current statSync(normalizedFile).mtimeMs against the stored value. If currentMtime > prev.read_mtime, the file changed — delete the stale entry and fall through to the normal first-read path (anatomy hint, no warning). If mtime is unchanged, the existing repeated-read warning fires as before.

src/hooks/post-write.ts

• Problem: After Claude writes to a file, the files_read session entry for that file still held the old read_mtime. The next pre-read of the same file would have to go through a statSync comparison to detect the change, which is fine — but the write hook already knows the file changed.
• Change: In section 3 (session tracker), add delete session.files_read[normalizedFile] immediately after updating files_written. This eagerly invalidates the read cache for the written file, so the very next re-read is treated as fresh without needing to compare mtimes at all.

---

Acceptance criteria

• Claude reads foo.ts → Claude writes to foo.ts → Claude reads foo.ts again → no repeated-read warning, anatomy hint re-fires
• Claude reads foo.ts → user edits foo.ts externally → Claude reads foo.ts again → no repeated-read warning (mtime changed)
• Claude reads foo.ts → nothing changes → Claude reads foo.ts again → repeated-read warning fires (correct behaviour preserved)
• repeated_reads_warned counter is not incremented for legitimate re-reads
• statSync errors (deleted file, permission denied) do not crash the hook — they fall back to the original warn-on-reread behaviour

How to test

1. openwolf init in a test project, open Claude Code
2. Ask Claude to read any source file — note the anatomy hit in stderr
3. Ask Claude to edit that same file, then read it again
4. Expected: no ⚡ OpenWolf: ... was already read warning on the second read
5. Edge case: ask Claude to read the same file twice without any write between — expected: warning fires on the second read (regression check)
6. Edge case: externally touch src/foo.ts between two reads — expected: second read is treated as fresh

Notes

• No schema changes
• No new dependencies added
• No server/API changes
• No breaking changes to existing consumers — read_mtime is an optional field; sessions without it (e.g. old _session.json files) fall back to the original behaviour because the if (prev.read_mtime !== undefined) guard is skipped
• statSync is already used in this file for the symbol-hint freshness check, so the import is not new
• The post-write.ts change adds files_read to the SessionData interface there (was only in pre-read.ts's local interface) to make the delete type-safe