[tommy-gun]

Follow-up to #24. That PR fixed collection discovery for issue #23, but introduced a regression that showed up on the first live capture.

The problem

Discovery walked the object graph depth-first against a global visit budget (maxVisitedObjects: 2000). Hitting the budget set #discoveryStopped, which ended the entire scan — not just that branch. So the first object-valued own property of stateObj could consume the whole budget and starve every sibling enumerated after it.

#24 removed the guard that had been shielding the old implementation from this:

if (!obj || typeof obj !== 'object' || Array.isArray(obj)) { return; }

The pre-#24 walk never entered arrays, so bulk data never cost anything. Once arrays and Map/Set values became traversable — which is exactly what issue #23 needed — nothing replaced that protection.

Observed on rss-feed (_store holds a Map of 27 feed sources, each with an array of parsed items): only _store.* got instrumented. _indicator.*, _poller.* and _repository.* were never reached, and because the calling functions were missing, every event collapsed to depth: 0 — the call tree, and with it the Self-vs-Total distinction, was gone.

Reproduced synthetically at realistic proportions (27 sources × 40 items, default limits):

	patchedFunctions	_store	_indicator	_poller	_repository
depth-first	26	✅	❌	❌	❌
breadth-first	37	✅	✅	✅	✅

The fix

Walk in level order. Methods sit near the root while bulk data (cached items, parsed entries) sits deeper, so breadth-first spends the budget on shallow objects first and truncates the deep data instead. truncated is still true in both rows above — the difference is what gets cut.

• #patchTree is replaced by #patchGraph (level-order queue) + #visitObject (single-node work). The four walk methods no longer recurse; they receive an enqueue callback.
• Objects are claimed when queued, not when processed, so an object referenced by several parents is never scheduled twice.
• Exhausting maxVisitedObjects now only ends the loop and marks the scan truncated. The hard #discoveryStopped flag stays reserved for maxPatchedFunctions, where a global stop is the intended behaviour.

No changes to #patchObject or the restore path — wrapper installation and descriptor restoration are untouched.

Notes for the reviewer

• testDataHeavySiblingDoesNotStarvePeers mirrors the rss-feed shape with maxVisitedObjects: 30 so the budget genuinely runs out. I verified it fails against the depth-first implementation — otherwise it would not be catching anything.
• One thing from the pre-Fix collection module profiling #24 captures will not come back, deliberately: _store._sources.values. The old code shadow-patched Map.prototype.values onto the instance, which measures the GJS runtime rather than the extension.
• Full suite green: stop hook exit 0 (ruff, eslint, mypy, syntax, WSL headless, pytest 59 passed), GJS bridge suite 9/9, inspector suite OK.
• Still needs live verification on GNOME. The proof here is synthetic — a fresh rss-feed capture should now contain _indicator.*, _poller.*, _repository.* and, importantly, non-zero depth values.

🤖 Generated with Claude Code