fix: make OpenRouter debug-fetch patch concurrency-safe

[Greg Land (bikeusaland)]

Summary

Fixes a concurrency hazard where the OpenRouter debug-fetch patch could leak permanently or be lost when two OpenWiki runs overlap in the same process (#411).

installOpenRouterDebugFetch swapped globalThis.fetch for the duration of a run and restored it blindly in a finally. With two overlapping runOpenWikiAgent calls — subagents, parallel tests, or a long-lived host process — the install/restore steps corrupted each other:

1. Run A installs, capturing the real fetch as its original.
2. Run B installs, capturing A's patched fetch as its "original".
3. Run A restores the real fetch.
4. Run B restores A's patched fetch — now leaked into the global forever.

Both runs also shared a single failure slot, so one run's captured error could overwrite another's.

Why not a per-model configuration.fetch hook

The issue suggested attaching diagnostics per-model like the Codex/Vertex clients do. That isn't available here: ChatOpenRouter (from @langchain/openrouter) extends BaseChatModel and calls globalThis.fetch directly — its constructor exposes no fetch/configuration field. A global wrapper is genuinely the only interception point, which is why the original code patched the global.

Fix

Make the global wrapper reference-counted:

• Install the wrapper once, capturing the genuine original fetch.
• Register a per-run failure sink so concurrent runs never share or clobber each other's captured failure. getLastFailure/clearLastFailure operate on the calling run's own sink, and debug events route to each run's own onEvent.
• Restore the global only when the last run detaches, reverting to the true original — never to a wrapper installed by a concurrent run.
• Redundant restore() calls on the same handle are idempotent (can't prematurely restore while another run is active).

The OpenRouterFetchCapture return contract (getLastFailure, clearLastFailure, restore) is unchanged, so callers are untouched.

Tests

New test/openrouter-debug-fetch.test.ts (4 tests): single-run restore returns the exact original; overlapping runs each capture their own 429 while restore stays reference-counted; non-OpenRouter requests pass through untouched; redundant restore is a no-op that doesn't disturb an active run.

Full suite green: typecheck, eslint ., prettier --check, and all tests pass.

Closes #411

🤖 Generated with Claude Code