Zwangsrüümig (role-lane-eviction)
When a census says "the role pool is live everywhere, no knob frees it" (the dead-knob
wall), the remaining move is to stop pooling: rewrite every site that uses the lane
with a construction that uses lanes provably free at that point, then drop the lane
from the allocator (alloc_n(3) -> alloc_n(2)), and make any residual use a build-time
panic (Toggle::Evicted panics on use).
How jackylee0424 did it (ecdsa.fail a6098537)
- Split by regime first. No-payload traversals (T2/T3): the evicted role binds to boundary-clean payload words (P0/P1) and a second scratch loan — cheap, even a T win (-28.6M emitted T there). Payload traversals (T1/T4): 13+ named rewrite families (witness-free records, FF/FFSW/PAIR forms, E5 prepared compare, K2C/K2D ladders, INV3 under a digit-sign premise, exact full-complement sign queries, hosted maps with schedule bounds, endpoint hosts on named Q lanes, ...).
- Every host must be proved free by layout or schedule, never observed-zero. The candidate host that was "refuted for c146-202" is the discipline: a host that is merely usually zero kills the row.
- Pay for it honestly. +8.8% executed T for -1 qubit. On a q x T board this is a Pareto move only where the exchange rate (T/qubit at the margin) favors width.
- Schedule-bound shortcuts: hosted maps carry a "counter < 64" schedule bound; nonfixed middles replace fixed middles where the fixed form needs a third host.
- Bake the switches. Each stage gate was an env knob baked on in mod.rs with an off-switch for A/B (C658_T23, C658_WP1..5, C658_DROP).
When a census "kill" is not dead
A census of the Q658 tree had proved every register knob dead (the pool needs its lanes simultaneously). The eviction answer routed around the census entirely: not "free a bit from the pool" but "delete the pool". Rule: a knob-level kill does not bound a site-level rewrite program. When you park a width, say which of the two you measured.