fix: sub-classify the dominant patch 'other' error bucket (Closes #1537)

[Lexus2016]

Summary

Automated evolution PR for issue #1537 — the patch tool's other error bucket was 81% of all patch failures (70/86 in a 7d introspection window): opaque to the digest and giving the agent no recovery signal, driving retry spirals up to 6 deep.

Root cause (verified against the real code path)

The fuzzy matcher's dominant failure message — "Could not find a match for old_string in the file" (tools/fuzzy_match.py:197, fired on every fuzzy-match miss) — matched neither classifier:

• Introspection (scripts/introspection_extract.py): _NO_MATCH_RE requires the literal phrase "no match", but the message says "Could not find a match" → fell to the opaque other bucket (the measured 81%).
• Runtime (tools/file_operations.py::classify_file_error): the fuzzy_match branch checked "did not match" / "no match" / ("match" + "block") — none match "Could not find a match" → fell to the generic error class with the unhelpful recovery "The operation failed. ... CHANGE the call.", giving the model no signal to re-read the file.

The recovery directive mechanism (hard refuse + write_file at threshold 3) already existed (#507/#1037) — the gap was purely the misclassification, plus surfacing the write_file escape one step earlier.

Changes (3 targeted fixes)

1. Introspection (scripts/introspection_extract.py): add a patch-no-match sub-category regex for "could not find (a) match", placed before _NO_MATCH_RE. The 70 previously-other failures now become a trackable, actionable signal in the digest.

2. Runtime classifier (tools/file_operations.py::classify_file_error): extend the fuzzy_match branch to recognize "could not find" / "not find a match", so the model gets the "Re-read the file and copy the EXACT current text" recovery directive instead of the generic one.

3. Earlier write_file nudge (tools/file_tools.py): on the 2nd consecutive no-match failure on the same file, surface write_file as the recommended escape BEFORE the hard 3rd-failure refuse-gate fires. The existing gate already refuses at threshold 3; this biases the model toward the reliably-terminating rewrite path one step earlier. Deliberately NOT gated on has_diagnostic — the consecutive-failure count + write_file signal is complementary to (not mutually exclusive with) the structured diagnostic.

Tests

• test_patch_no_match_classified_not_other — introspection classifies "Could not find a match" as patch-no-match, not other.
• test_could_not_find_match_is_fuzzy — runtime classifier routes it to fuzzy_match recovery.
• test_second_consecutive_no_match_nudges_write_file — the [FIX] patch tool 'other' error bucket still dominant — 70/86 failures (81%) after decomposition fix #1537 nudge fires on attempt 2 (not 1), recommends write_file, and is not the hard refuse-gate.
• test_first_two_failures_no_hard_escalation (existing, updated) — tightened to assert on PERMANENT FAILURE so the new soft nudge remains valid.

Validation

• Pre-PR targeted test shard: ✅ PASS (3 shards)
• ruff check: ✅ All checks passed
• Full related suites (324 tests across patch/file/fuzzy/introspection/recovery): ✅ all pass
• Diff: 141 insertions, 3 deletions across 6 files — within the 200-line self-merge cap.

Closes #1537