#2777 deleted a recorded closed negative along with its stale preamble

Row: GFX1100-TG200

#2777 removes this paragraph from
docs/bench-evidence/gfx1100-tg200-t2b-20260823.md:

## Session-state note (hindsight unavailable)

Hindsight memory store was returning 500s during this session, so the
campaign state is recorded here instead: T2b commit is 90f7ca064; prior
levers are 369e4b044 (merged gate_up) and 69c514a1d (quant-cache rejection);
the pointer-keyed activation-quant cache approach is UNSOUND under the
DevicePool block-recycling allocator -- do not retry it without a
content-identity signal.

The preamble is stale and worth dropping. The last sentence is a closed
negative, and it is the only place the tree records it: a recycled DevicePool
block can present a pointer the cache still believes it owns, so the approach
cannot be made correct without a content-identity signal. A recorded negative is
what stops the next session re-running the same experiment, and AGENTS.md is
explicit that evidence is not deleted to reduce context.

Fixed in the same flow: the finding is restated without the preamble.

Start with docs/bench-evidence/gfx1100-tg200-t2b-20260823.md and read the evidence-preservation guidance in AGENTS.md. Confirm that the stale session-state preamble is removed while the closed negative about pointer-keyed caching and recycled DevicePool blocks remains recorded in the restated finding.