chore: update drain3 default log pattern weights

[github-actions]

chore: update drain3 default weights from weekly training run

Summary

Refreshes pkg/agentdrain/data/default_weights.json with the latest weekly
drain3 training run. No Go source or test files were modified.

---

What changed

New error clusters (ids 7–10)

id	Template / description	type
7	web-fetch MCP tool not available	missing_tool
8	mcpscripts-gh CLI not mounted; only github_pr_query / github_issue_query / github_discussion_query available	missing_tool
9	web-fetch tool not available → direct web-fetch smoke step could not execute	missing_tool
10	/tmp/gh-aw/comment-memory directory absent; no markdown files to append haiku to	missing_data (data_type=comment_memory_files)

next_id in the error stage bumped 7 → 11 to accommodate the four new clusters.

Updated existing error clusters (ids 1–6)

id	Before	After
1	playwright-cli not on PATH	reason=missing tool/permission issue: numerous permission denied errors (size 1 → 2)
2	web-fetch Codex toolset template	Detailed close_discussion safe-output not available error
3	Cache-memory miss (multi-field template, size 5)	Simplified: data_type=cache_memory reason=cache_memory_miss type=missing_data (size 5 → 3)
4	close_discussion safe-output missing-tool template	reason=security tool= type=missing_tool
5	Security/tool template (size 2)	Detailed mcpscripts-gh unknown-tool error (JSON-RPC -32602) (size 2 → 1)
6	Security / gh CLI missing	web-fetch MCP tool is not available in this runtime

Cluster size updates

Stage	Cluster	Old size	New size
finish	—	373	500
plan	—	136	449

tool_result clusters

Multiple cluster templates updated to reflect newly observed patterns (e.g. PR-nudger run summaries).

JSON formatting

ExcludeFields arrays and template arrays reformatted from single-line to multi-line style throughout the file for readability.

---

Why

The drain3 weight file is regenerated on a weekly cadence from real agent run logs. This update:

1. Captures four new recurring error patterns seen in production runs — primarily missing_tool errors for web-fetch and mcpscripts-gh, and a missing_data error for the comment-memory directory.
2. Refines existing cluster templates to better match observed log lines, reducing mis-classification noise.
3. Reflects real cluster growth in the finish and plan stages (size 373→500 and 136→449 respectively), keeping weight scores calibrated to current traffic volume.

---

Testing

Data-file-only change. No code paths altered; existing drain3 unit tests cover the weight-loading logic.

---

Files changed

File	Description
pkg/agentdrain/data/default_weights.json	Weekly drain3 weight refresh

Generated by PR Description Updater for issue #36190 · sonnet46 750.6K · ◷