04: shared-logic sweep part 2 — streaming and tool-call edges

[elkaix]

Part of #341.

What to build: bounded shared-logic sweep, part 2 — compare streaming/tool-call edge handling (partial tool-call input, interrupted streams, tool-call summaries) between the private reference checkout and our agent surfaces. Land only proven defects with failing-first regression tests; ledger every examined subsystem.

Blocked by: #3 (sequential sweep: shared agent-core-v2 surface, one writer).

Status: ready-for-agent

• Comparison notes written to the private ledger with decisions
• Every landed fix has a focused test failing pre-fix, passing post-fix
• Touched packages' suites green
• Leak gate clean

Demo path: red-then-green evidence per fix.

Capability: deep/medium
Claims: exclusive: packages/agent-core-v2/src/, apps/pythinker-code/src/

[elkaix]

Comparison-only outcome: no proven defect; decisions recorded in the run's private sweep notes (partial-input preview, interrupted-stream summaries, block ordering all already-handled or not-applicable with evidence). One uncited lead (gateway onToolCallDelta frame behavior) recorded for a future run.