srv1 runner width subtracts the served-slice charge (14 -> 11); shakedown index derived

[gunbai-bot]

Operator option A (2026-09-26): srv1's runner width subtracts the served-deployment slice charge.

What changed

• gunbc.runner_slot_allocation envelope: subtracts gunbc.live_deploy.served_slice_charge host_served_deployment_claim (new envelope_after_control_plane), using the control plane's arms: NotPlaced is a grounded zero, and Unmeasured refuses the envelope. Before this, the width ignored a charge that the verdict (gunbc.ci_runner_placement host_standing_at_width) counts. That gave two answers to one pool.
• Second fork, exposed once the first closed: the verdict charged every slot at the fleet row, including the shakedown slot, which the width charges its full cell + controller. The new gunbc_runner_pool_for_host_at_width is read by the placement verdict and by the conservation witness.
• The shakedown slot index is derived (parent ruling from proud-deer-538): it is the last index under gunbc_runner_committed_width instead of the authored 13. The width change silently falsified the old note ("below its memory-admitted width"). The host is split into gunbc_microvm_shakedown_host, so the width does not depend on the index. New claim the_shakedown_index_is_the_top_of_srv1s_committed_width.

Evidence (executed, local claim_batch, private target dir, base f17c19a vs head)

• srv1 committed width: 14 on base, 11 at head (probe claims via gunbc_runner_committed_width / host_memory_admitted_width).
• committed_width_fits_on_every_host_with_an_established_charge: FAIL on base (discriminating RED), PASS at head.
• one_more_slot_than_memory_admits_refuses_on_every_host (+1 control): PASS at head. It reddened after the first fix alone, which is how the shakedown under-charge surfaced.
• srv1's placement standing: Refused (oversubscribed) on base, Derived at head.
• Census: every test fn in the 42 witness files importing runner_slot_allocation / ci_runner_placement / runner_host_deploy / fleet_host_budget / runner_microvm_slot_controller / runner_lifecycle. 384 base vs 384 head identities: 0 PASS→FAIL, 9 FAIL→PASS. The 9 are the conservation witness plus 8 in host_converge_slice1_witness_test, which were red on main because srv1 had dropped out of the converge policy. Caveat: floor_cold_build_receipt_witness_test stopped the batch (wet witness), so the census covers the files around it, same population on both sides.
• Pre-existing and not touched: saturation_refuses_rather_than_reporting_the_ceiling FAILs on base and head.

Enrolment

The conservation witness is in no CI lane. It should be enrolled: it's cheap (2,810 eval_steps), and it is the only check that sees width vs. verdict disagreement. But enrolling the file surfaces the pre-existing saturation_refuses… red, so enrolment needs that red resolved or declared first. Left to a follow-up.

srv1 observation (by proud-deer-538, 2026-09-26, not re-measured here)

502 GiB physical and 12 runners active at 26 GiB caps (free -g). Roadmap slice peak 51.1 GB vs 52.6 GB MemoryHigh (systemctl show -p MemoryHigh -p MemoryPeak <slice>).

Apply consequence: NOT RUN

Width 14→11 makes srv1-12 and srv1-14 required retirees (via the open runner-slot-26-25-gib transition, prior_width 50, required_retirees_for_host), and the shakedown cell renames srv1-13 → srv1-11 (fabric-cell-srv1-11, gunbc-microvm-slot@srv1-11). srv1-11 is live as a GitHub slot today. Per proud-deer-538, the plan's REMOVE/rename lines go to them before any srv1 apply.

🤖 Generated with Claude Code

[gunbai-bot]

Closing: superseded by #12339 (96e6331), which landed the same served-charge subtraction, charges the shakedown slot at real size, and sets srv1 width 11, keeping srv1-13 by identity under an operator ruling. #12357 is rebased to carry only the host_allocation_conservation gate enrolment that main still lacks. — sent from proud-deer-538