The 2026-09-23 floor sizing receipt: uncensored at the 26 GiB row, the microVM shape resolves, the fit stall retires

[gunbai-bot]

Deliverable 4 of the floor memory attribution lane (work item adhoc-80d33957-f14). This builds on #12078 (instrument) and #12088 (the one reduction). Until those land, the diff also shows their commits; the change here is the last commit.

The receipt

gunbc.floor_demand gunbc_floor_memory_stat_receipt_2026_09_23 is a dated successor to _2026_09_21, and it is now the input to gunbc_floor_held_set_standing.

• Source: required-floor run 35815214135, job 107035188664, actions-runner@srv4-14.service.
• Base measured: it is a COMBINED base, and the receipt says so, so neither change is credited with the other's share: main after floor memory attribution baseline, one reduction, refreshed sizing receipt #12081 (one corpus read for both prepares), plus The CI diff policy is its own module, so the floor's diff observer stops resolving the whole CI spec #12088 (the diff observer no longer resolves the whole CI spec), with Floor memory instrument: a beat at every phase seam, with the leaf's swap, PSI, local events and charged processes #12078's instrument.
• Rows: all 40 stat beats the run printed, generated from the log's own lines rather than retyped, one at every phase seam plus the watchdog minutes, each with the leaf's own memory.swap.current. beat is the sample ordinal, and the instrument's when= token is the comment above each row.
• What it does not establish: it is one warm, repeated run on the 26 GiB host-slot row. It is not a cold attempt (that accounting stays with gunbc.floor_cold_build_receipt), not the 28 GiB shakedown slot, and not a fresh attempt cgroup, which is why it carries memory.stat and never the leaf's memory.peak. Swap policy: SwapMaxDeclaredUnread (32 GiB, gunbc.runner_slot_allocation gunbc_runner_slot_desired memory_swap_max).

Carrier and folds

• FloorSeam gains one arm per seam the instrument now names (lane-roster through publication). receipt_peak_seam answers each; the match is exhaustive.
• FloorMemoryStatBeat.swap: BeatSwap is SwapRead { bytes } or SwapUnread. Every earlier beat carries SwapUnread, never a fabricated zero.
• receipt_peak_reading_standing reads the peak beat's leaf swap first. A zero reading means none of the leaf's pages were in swap at the peak, so a slot that COULD swap did not at that instant, and the peak is PeakIsUncensoredSampledMaximum. A non-zero reading censors as SwapHeldPagesAtPeak { beat, swapped }. SwapUnread falls through to the declared-policy rule, unchanged, now peak_standing_from_swap_policy. This is the consumption floor_receipt_uncarried_observation_frontier asks for, and that frontier's text is updated: swap leaves it; PSI, oom and process overlap are now printed every beat but not yet carried.

The fit, re-judged against current demand

With the new standing, gunbc_runner_microvm_shape resolves on production inputs, with the cell row (26 GiB) and the realization allowance (1 GiB) unchanged and nothing deleted or supplied.

• The flip: the expecting-red probe the_production_shape_refuses_today_on_the_memory_remainder_with_all_three_figures flips (DESIGN §4b(4)) to the regression control the_production_shape_resolves_on_the_memory_remainder_since_the_2026_09_23_receipt. It cites the DEMAND RECEIPT as what moved and derives the sum through the same folds, with no transcribed figures.
• The stall row: gunbc.guarantee_stall.runner_microvm_floor_fit_stall, whose next_rung_trigger is exactly this capability, is deleted along with its roster entry. Its executed evidence stays enrolled. the_production_shape_answers_according_to_the_floor_demand_standing (either-state) asserts the refusal triple the day a later receipt moves demand back over, and the supplied-input controls keep ShapeRefusedWorkMemoryExceedsCellRemainder executed.
• Prose updated: the rung drop slot_row_pinned_below_demonstrated_demand_unrefused keeps its own trigger, and only its mention of the stall changes. docs/plans/microvm-and-floor-wind-down-state.md §2 gains a dated note and keeps the prior state as history.

Margin, stated as a bet (§4d): the demand is one warm run's SAMPLED maximum, a lower bound on true demand at 60 s cadence plus seam beats, and the sum fits by about a gibibyte. A later receipt can put it back over, and the either-state witness is what would say so.

New witnesses

• a_leaf_swap_reading_at_the_peak_decides_before_the_declared_policy: one declared-swap receipt driven three ways (read zero → uncensored, read non-zero → SwapHeldPagesAtPeak, unread → policy cause).
• the_2026_09_23_receipt_reads_an_uncensored_sampled_maximum_through_its_leaf_swap: real rows, asserting the ROUTE (peak beat swap read zero, and the policy alone would censor), not only the answer.
• the_standing_receipt_attributes_its_peak_to_a_named_seam: re-pointed to the 09-23 standing (peak during publication), still asserting the 09-21 attribution.

Executed

claim_batch over dag/test/claim/floor/floor_demand_witness_test.dag (28) and dag/test/claim/runner/runner_microvm_witness_test.dag (48), on BuildBuddy with GUNBC_MEMORY_BUDGET_BYTES and EstimatedMemory=48GB: exit 0, which claim_batch returns only when no claim failed, so 76/76 passed. The required CI check runs no claims since #11742, so this is the execution.

Not done

The 28 GiB fresh-attempt-cgroup run. It waits on #12011's shakedown slot, and the slot lane has not answered. This receipt judges only the 26 GiB row.

🤖 Generated with Claude Code