Bound what one guest can make the chat hold or compute (review, area 2)

[adamjohnwright]

These come from the max-level review of the chat flow. Each was reproduced by the reviewer.

Finding	Fix	Checked by
The typed-yes regex backtracked quadratically on the raw message. One offer, then "y" + 990K spaces, would stall the shared event loop for about 90 minutes.	Messages over 40 characters are not a yes, and the pattern uses possessive quantifiers	Timing test (under 0.1s). Under sabotage the test hangs until killed.
Chat messages had no length cap. A 1M-character message was kept in the checkpoints (3 MB) and sent to the model on every later turn.	Messages over 60K are declined; questions over 8K are declined before reaching the model	Browser: an oversized question gets the polite refusal, and the next question is answered
Chainlit keeps every message in chat_contexts[session] forever, including rate-limited ones.	Dropped after each message and when the session ends (nothing here reads it)	Browser: multi-turn chat still works
Guest graph threads, offers and seed state were never freed.	Deleted on on_chat_end for guests, who cannot resume	Store test
Gene-list seeds and offers kept the whole 60K message and every identifier, about 0.85 MB per offer.	Capped at 8K of text and 3,000 identifiers	—

The GSA validator problems (unbounded work on the event loop, encoding, R headers, CSV) will be in the GSA PR.

./checks.sh passes. The gene-list browser suite passes.

🤖 Generated with Claude Code