Fix what the model is told (review, area 1a)

[adamjohnwright]

This is the third group of fixes from the max-level code review of src/api and src/handoff.

Finding	Fix	Test (each sabotaged and seen to fail)
A handoff seeded the summary's data without the rules the summary was written under, so follow-ups were answered from input known to produce "12 of 1280".	summary_instruction() and DATA_RULES live in summarise.py. The endpoint and the seed both use them.	Seed carries the rules
"Every one of them is significant" was added whenever the count was inexact, including when only 2 of 12 were significant.	Said only when every shown pathway is significant. Otherwise UNKNOWN_COUNT_INSTRUCTION.	12/12, 2/12 and 0/12 cases
A temporary failure at claim time told the model the result was deleted.	The seed is given the outcome and says "temporary" for a failure.	Failed vs gone
Summaries ignored config.yml's llm: block and the temperature rules, so fixed-temperature models failed every summary. Content blocks were dropped, and the stream still reported summarised.	_summary_llm() resolves the model the way the chat does. _text() reads content blocks. An empty stream is a failure.	o3's temperature; blocks read; empty stream not summarised
Two concurrent requests generated different summaries, and the later one overwrote the earlier. The aggregate fallback regenerated instead of reusing.	Single-flight per key. The store never overwrites. The fallback reuses a stored aggregate.	Two readers, one model call, one text; store keeps the first

Verified

• ./checks.sh passes.
• The browser test passes against a local instance: an analysis summary from the real endpoint, a handoff, the same summary shown, a follow-up answered from the data, and an unknown id reported.

Also removes new_thread_id, which nothing called.

🤖 Generated with Claude Code