enhancement(apply_diff): fuzzy-match success warning, file-changed signal, large-file perf guard (trial agent feedback, epic #1375)

[easonLiangWorldedtech]

Observed by the trial agent (epic #1375 trial build, #1413) while running a parallel-writer read-modify-write experiment on apply_diff (two writers on one file + a looping anchored edit).

Working well (no action needed)

1. The mismatch error payload (Best Match Found + Original Content) gives the agent everything it needs to re-anchor in a single round-trip - no extra read_file required.
2. The fuzzy threshold correctly failed loud instead of silently mis-hitting (LOOP_70 vs LOOP_72 under a 92% threshold). Default exact (1.0) + diffFuzzyThreshold opt-in + clamp to [0.5, 1.0] is the right shape.
3. Atomicity held: a two-block diff with one bad block failed as a whole - no partial application.

Suggested follow-ups

1. Fuzzy success warning. A successful match below 100% similarity is applied silently today. Suggestion: include one line in the success result, e.g. "applied at 94% similarity (anchor line 70 -> actual line 72)", so the model can decide to verify.
2. Explicit file-changed signal. Today the agent must infer a concurrent writer from the Original Content payload. An explicit mtime/hash-comparison flag in the error (e.g. "file changed since last read: true") lets the agent pick the right retry strategy: "my diff is wrong" vs "file changed externally, re-anchor".
3. Performance guard for large files. The sliding-window + per-block Levenshtein is O(n*m); on multi-thousand-line files with large search blocks this may be noticeably slow. Suggestion: a window cap, or an exact quick-pass before the fuzzy pass.
4. Document the all-or-nothing block guarantee in the tool/README copy (it is the right behavior for a "surgical edits" tool).

Context

• Epic: [EPIC] File Write Safety Prevent Concurrent Write Races Data Corruption #1375 (file-write-safety series). Trial: feat(fws): file-write safety series — trial build (all 15 component PRs) #1413.
• The two change-card gaps found in the same trial session (apply_diff missing the per-step change card / restore button; missing open-file button) were fixed in the component PRs (feat(checkpoints): per-step change cards and changeCardDetail setting (B3a, #1375) #1411 / feat(webview): change cards UI and rollback buttons (B3b, #1375) #1412); this issue tracks only the fuzzy-matching observability and performance suggestions above.