False-positive [bio] safety block sticks for rest of session — technical physics discussion

Environment: Claude Code desktop app, model claude-sonnet-5

A session doing routine ice-physics/aerosol modelling work (discussing the Feff collision-angle parameter for blowing-snow-sourced sea-salt aerosol generation vs. in-cloud collisions, sin^2(theta) probability calculations, manuscript editing) started returning this error on every subsequent turn:

API Error: Sonnet 5 can't help with this. Start a new session to continue.
Learn more: https://www.anthropic.com/legal/aup
Details: `[bio]`

The first blocked turn came right after the user asked the assistant to confirm/improve a statement distinguishing Feff=0.18 (25 degrees, clouds) from Feff=0.036 (11 degrees, blowing snow) -- a purely physical/geometric probability discussion, no biological content. All subsequent turns in the session were also blocked with the same [bio] tag, including a plain "are you there?" -- suggesting the block is sticky for the rest of the session once tripped, rather than re-evaluating each turn's actual content.

This is the second occurrence of the same failure mode I've hit; the first is reported separately (session local_61cd610f-9038-4299-a671-36f96e5559f8).

Session ID: local_c9c8b339-3c85-4d93-9909-2ba0a263d76d
First blocked Request ID: req_011CeuYV8wUaFpmgTrzrw9uR
Subsequent blocked Request IDs:
req_011CeuYaVMT4RsFkeKGx7SCN, req_011CeuYcxZcjLxaGJEuHWoKz

Reporting as a likely false positive -- no biological content is present anywhere in the visible transcript around the trigger point.

Reproduce the false-positive block in the Claude Code desktop app using session local_c9c8b339-3c85-4d93-9909-2ba0a263d76d and the listed request IDs, then compare the first blocked turn with later benign turns. No source file or test is named; done means a false-positive [bio] trigger no longer blocks subsequent unrelated turns for the rest of the session.

索引モデルが issue の本文から書いたものです。