feat(langgraph): emit a progress milestone when an oversized subtopic batch is rejected

Summary

When research_subtopic rejects a batch above the hard limit (>10 subtopics), it returns the error text to the lead agent and logs a warning, but never calls progress_callback — so progress-UI consumers never see that a batch was rejected and why.

Field	Value
Severity	suggested (review section was Should-fix; downgraded on promotion — the review itself frames this as "design choice, not a defect", impact is observability-only)
Origin	Review of #5390 (post-merge extraction)
Review section	Should-fix
Classification	PR-introduced (rejection path added in #5390)
Verified at	main@99c07fe97fca
Dedup	distinct · 0 hits reviewed

Trigger

Invoke research_subtopic with 11+ subtopics: the requested_count > MAX_SUBTOPICS_HARD_LIMIT branch logs logger.warning and returns the rejection string before any progress_callback emission; the queued-overflow path further down does emit progress + metadata, so the two paths are observably inconsistent.

Impact

Research-progress UIs show nothing for a rejected batch — the run appears to silently skip a step; only the lead agent's text and container logs carry the reason.

Pinned evidence

• 99c07fe97fca:src/local_deep_research/advanced_search_system/strategies/langgraph_agent_strategy.py:L629-L641 — hard-limit rejection: warning + return, no progress emission
• 99c07fe97fca:src/local_deep_research/advanced_search_system/strategies/langgraph_agent_strategy.py:L728-L746 — the overflow path that does call progress_callback (and attaches overflow metadata)

Fix direction

Emit a milestone before returning, e.g. progress_callback(f"Rejected batch of {requested_count} subtopics (hard limit {MAX_SUBTOPICS_HARD_LIMIT})", None, {"overflow_strategy": "rejected"}), and relax/extend the rejection tests to expect it.

Targeted tests

Add/extend a case in TestResearchSubtopicToolOverflow: invoke with 11 subtopics, assert the captured progress_callback received the rejection milestone and the return text still carries the split-into-batches guidance.

Provenance

Extracted from review of PR #5390 (post-merge extraction).