mtcollins1 one-socket runner offer from census meminfo (#13252 re-derived)

[gunbai-bot]

Summary

• Re-derives gunbc#13252 on current main: gunbc.ci_fleet mtcollins1_offer at the measured one-socket shape (one Q80-30-shaped processor, 80 threads; empty memory list), not enrolled in fleet_owned_compute_offer_catalog.
• Adds gunbc.host_memory_observation census_meminfo_reading / host_usable_ram_from_capture / host_fixed_overhead_from_capture. Slot count is gunbc.runner_slot_allocation committed_width_of over that reading and gunbc.runner_slot_desired (22 GiB / 21 GiB / swap 0). Production gunbc_ci_host_usable_ram_for(mtcollins1) stays HostUsableRamUnmeasured until a wet census receipt exists.
• gunbc.session_residency classifies mtcollins1 as SessionsOperatorDeclaredAbsent (operator ruling 2026-10-03). operator_declared_absent_host_count now walks session_residency_evidence_hosts so this override is in the census (review 4232891432 on mtcollins1 runner offer at measured one-socket shape + census meminfo reader #13252).
• Renames ResourceUnread.SectionAbsent → ResourceSectionAbsent so the census reader can share a closure with HostCaptureSectionReading.SectionAbsent (resolver refuses duplicate variant names).

History, not a base: session/fierce-hawk-431-mtc1-offer / gunbc#13252 (closed in the v1 closeout). Floor-runner PRs #13211, #13288, #13640 were read only for host facts (one-socket 80 CPUs).

Deliberately not carried

• RAM-backed tmpfs / HostWorkspaceStanding / WidthWorkspacePeakUnestablished / moving selected_slot (floor-runner qualification; dead after the v1 withdrawal). That also makes review 4232891418 (P1: filter qualification runs by full subject) inapplicable — those rows are not in this derivation.
• Enrolment into fleet_intent_offers and a hand-copied MemTotal row. Width is derived from a supplied census receipt; a wet meminfo is still the enrolment trigger (operator-only census boot).
• Stacked CPU-axis commits from CPU admission axis reads each host's offer, not one catalog row #13197: already on main.

Conformance

• Fleet/compute: gunbc.ci_fleet ComputeOffer, gunbc.runner_slot_allocation committed_width_of, gunbc.runner_slot_desired.
• Observability: census meminfo via host_capture_section_reading + observe_memory.
• Leasing/sessions: gunbc.session_residency / gunbc.session_reservation_policy.
• Every new declaration is consumed by test.claim.mtcollins1_runner_offer_witness and/or host_deployment_standing_given.

Evidence

• claim_batch --source-root dag --source-root src/v2 --entry dag/test/claim/runner/mtcollins1_runner_offer_witness_test.dag --functions census_meminfo_receipt_reads_as_measured_usable_ram,census_meminfo_RED_non_reads_refuse_by_cause,mtcollins1_RED_nominal_dram_total_reaches_neither_offer_nor_reader,mtcollins1_offer_cpu_width_equals_the_measured_one_socket_reading,mtcollins1_RED_both_sockets_disagrees_with_the_measured_reading,mtcollins1_carries_no_sessions_and_owes_no_reservation,mtcollins1_production_usable_ram_is_unmeasured_until_the_census_receipt,mtcollins1_placement_accepts_the_offer_over_supplied_one_socket_facts,mtcollins1_RED_placement_refuses_until_usable_ram_is_measured,mtcollins1_derived_width_moves_with_the_census_meminfo_and_the_slot_row,census_meminfo_receipt_reads_as_measured_fixed_overhead — exit 0, all 11 PASS (local seed /tmp/cargo-target-sunny-crab-860/release/claim_batch).
• claim_batch … --entry dag/test/claim/session_residency_witness_test.dag --functions managed_host_absence_is_declared_not_observed — exit 0, PASS.
• claim_batch … --entry dag/test/claim/machine_intake/host_resource_observation_witness_test.dag --functions an_absent_section_is_distinct_from_an_unparsed_one — exit 0, PASS.

Operator-only (not done here): one-socket census boot on mtcollins1 whose meminfo receipt this reader will turn into the production RAM/overhead rows.

Test plan

• Local claim_batch over the offer, residency census, and resource-unread rename
• Required witnesses job on this PR

Made with Cursor

[gunbai-bot]

Review 78456 (dashboard Claude request-changes on 154a959): agreed. The Q80-30 catalog bind was a bet marked only in a comment, which DESIGN §4d and §4c both refuse. Head now carries mtcollins1_socket0_cpu_part_standing as TranscribedUncited with a census read obligation (model number and Hertz off socket 0), and mtcollins1_socket0_cpu_part_is_established is false until that standing is CitedToAuthority or ObservedByReceipt. Thread width stays the measured 80. Claim mtcollins1_socket0_cpu_part_is_a_typed_bet_not_a_deduced_catalog_row holds.

CpuFacts still requires a catalog row to inhabit ComputeHost.processors; the standing is the typed warrant for that row, not a second comment.

[gunbai-bot]

Review 78464: deleted mtcollins1_socket0_cpu_part_is_established. The witness now asserts cited_figure_read_obligation(standing: mtcollins1_socket0_cpu_part_standing) is Present, which is the home accessor for TranscribedUncited.

[gunbai-bot]

Review 78456 (on 154a959): verified against current head 03c8794.

The finding was real on that commit: Q80-30 catalog + Hertz sat on mtcollins1_socket0_cpu with the bet only in a // block. Head now carries mtcollins1_socket0_cpu_part_standing as TranscribedUncited with a census read obligation (model number and Hertz off socket 0). cited_figure_read_obligation is Present; the witness mtcollins1_socket0_cpu_part_is_a_typed_bet_not_a_deduced_catalog_row asserts that. Width still uses compute_host_cpu_admitted_width → cpu_facts_threads (the measured 80). Hertz and model number are not an input to committed_width_of.

CpuFacts still requires a CpuModelCatalogRow; there is no lawful empty-frequency arm without forking that type. The catalog bind is the bet the standing types, not a deduced part. Enrolment stays out of fleet_owned_compute_offer_catalog until the census meminfo receipt and that standing is no longer TranscribedUncited, so priced selection cannot yet consume the Hertz as a fact.

Review 78471 approved this head after that change (and after deleting the extra boolean helper from review 78464).

— sent from sunny-crab-860

[gunbai-bot]

Review 78464 (on 2d59a77): verified against current head 03c8794.

The helper mtcollins1_socket0_cpu_part_is_established is gone (no remaining references in the tree). The witness mtcollins1_socket0_cpu_part_is_a_typed_bet_not_a_deduced_catalog_row asserts cited_figure_read_obligation(standing: mtcollins1_socket0_cpu_part_standing) is Present, which is the home accessor. That is commit 03c8794. Review 78471 approved this head after that deletion.

— sent from sunny-crab-860

[gunbai-bot]

Review 78471: the Hertz-as-fact caveat is the enrolment gate, not a leftover defect. mtcollins1_socket0_cpu_part_standing stays TranscribedUncited until the census reads the socket-0 model number; mtcollins1_offer stays out of fleet_owned_compute_offer_catalog until that standing is CitedToAuthority and the one-socket meminfo receipt exists. No CpuFacts fork in this PR — that type still requires a catalog row, and width already consumes only cpu_facts_threads (the measured 80).

— sent from sunny-crab-860