A handoff for the session that plants the probe, and the three numbers the murderboard caught in it

[syncytium2]

The app's sweep can now refuse a setting that cheats, and it will never do so — because the thing it refuses on is not in the data. That is the state of play the handoff leads with, and closing it is one change to the browser's generator plus the parity test it earns.

Filed in docs/handoffs/ not at the root: nothing in this session is half-done, it all landed. The root file belongs to the 2026-08-27 session and is still live by its own test — _still_live is any(state == "OPEN"), not all, and #292/#53/#50 are open. I assumed the opposite and checked before writing it down.

Murderboarded — 11 of 11 roles, 3 rounds, stopped on the severity floor

Run record: docs/reviews/2026-08-28-the-gate-is-in-the-app-and-inert_2026-08-28.md. Roster gate passes.

It earned its place on the first check — three factual claims were wrong, all authoritative-sounding:

claim	recomputed
suite 1,507	1,536 passed, 16 skipped, 1 xfailed — the draft's figure came from another worktree, pre-merge
"labSpec() consumed only by api.train"	two readers; paintLabWhat() displays it
"test_hooks_installed fails"	passes here — true only of a fresh clone

A fourth came from the blind pass and was introduced by a fix: the opening said "one job in one file" while a fix eleven lines below told the reader to add a parity test in a second file.

round	blocking	major	minor
1 · full review	2	3	3
2 · blind	0	1	1
3 · blind	0	0	2

The generalisable finding, written into the record: a status document written from a worktree quotes that worktree's world. Two of three blocking findings were measurements true where taken and false where they would be read.

Residual ⚠ — three, none silent

1. Role 2 ran single-pass, which the process forbids for attribution deliverables — this session runs under a standing instruction not to spawn agents. Exposure is low (the handoff restates an attribution settled in ADR-0002, it does not make one), but the rule exists because a single pass inherits the drafter's search history, and this one did.
2. F1-neutrality of planting the probe is a prediction, labelled as one, with the instruction to measure before and after.
3. Nobody was asked whether the user's simulated folder should carry a probe block — a product question the handoff raises and does not settle.

This review found and fixed 8 defects. It is not a correctness proof.

🤖 Generated with Claude Code