mtcollins1 runner offer at measured one-socket shape + census meminfo reader

[gunbai-bot]

Leg 4 of node adhoc-da1df2b9-c50, parts (a)+(b) plus the residency row. Stacked on #13197, so the diff includes its two commits until that lands. Draft until the operator's one-socket census boot produces a meminfo receipt.

What lands

• (a) Census reader: gunbc.host_memory_observation host_usable_ram_from_capture. It reads the capture envelope's meminfo section (via host_capture_section_reading) and parses it with observe_memory. The result is HostUsableRamMeasured (MemTotal × 1024 bytes), and the provenance names the machine and the run. Each kind of non-read refuses with its own cause: absent, pending, malformed, nonzero exit, unparsed. No mtcollins1 row is added to gunbc_ci_host_usable_ram_for; that row is the reader applied to the real receipt, so the wet number drops in with no new code.
• (b) Offer: gunbc.ci_fleet mtcollins1_offer. It has ONE processor, per the operator ruling of 2026-10-03 that admits mtcollins1 with socket 1 out. It lists no memory devices, so no nominal capacity enters. The part row (Q80-30) is recorded as a bet: it was chosen because its 80 threads equal the measured count.
• Residency: gunbc.session_residency gives mtcollins1 SessionsOperatorDeclaredAbsent, from the operator ruling of 2026-10-03 relayed by eager-gull-22. Dissolves on a persistent install or any session placement.
• Rename: ResourceUnread SectionAbsent → ResourceSectionAbsent. The reader is the first closure that joins it with HostCaptureSectionReading SectionAbsent, and the resolver refuses duplicate variants (measured).

Evidence (test.claim.mtcollins1_runner_offer_witness; all evaluated true via gunbc run at c683acf on BuildBuddy)

• census_meminfo_receipt_reads_as_measured_usable_ram: the inhabitance claim on the real reader route, with a supplied receipt.
• 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: checks against every one-socket boot that recorded a CPU reading in mtcollins1_socket1_investigation_boots.
• mtcollins1_RED_both_sockets_disagrees_with_the_measured_reading: 160 vs the measured 80.
• mtcollins1_carries_no_sessions_and_owes_no_reservation
• All 16 claims in test.claim.machine_intake.host_resource_observation_witness still hold after the rename.

Declared frontier (not in this PR)

mtcollins1_offer is not in fleet_intent_offers. Enrolment, and an end-to-end placement-accepts witness, wait on three things:

1. the census MemTotal row, read from the receipt by this reader;
2. a fixed-overhead row for mtcollins1;
3. the width chain reading a supplied facts record. That is the next PR, after CPU admission axis reads each host's offer, not one catalog row #13197, and it is also where the RAM-backed tmpfs workspace gets charged inside slot memory.

gunbc.managed_host has no measured-silicon field yet (it is cut 1 of docs/plans/managed-host-untangle.md), so this PR imports the mtcollins1 census modules directly in the witness. The offer will read through that seam once it exists.

🤖 Generated with Claude Code

[gunbai-bot]

Head b74b7c6 merges main (#13268's facts-record seam) and adds the end-to-end claims. All 9 claims in test.claim.mtcollins1_runner_offer_witness evaluated true on BuildBuddy at this head.

• mtcollins1_placement_accepts_the_offer_over_supplied_one_socket_facts: host_deployment_standing_given accepts mtcollins1_offer with derived runner_count=4. The record is mtcollins1's production host_width_facts, so session (the real residency row), control plane, served claims, disk and shakedown slot are all real. Three fields are supplied:
  • cpu: the offer's measured 80-thread width;
  • usable_ram: host_usable_ram_from_capture over a DESIGNED meminfo receipt (MemTotal 131000000 kB), the same route the wet receipt will take;
  • fixed overhead: a designed 8 GiB allowance, because mtcollins1 has no overhead row yet.
• mtcollins1_RED_placement_refuses_until_usable_ram_is_measured: the same record with today's real RAM standing (unmeasured) refuses UsableRamUnavailable.
  The width 4 is a function of the designed RAM and overhead and is not a prediction. Still a draft: it waits for the census receipt (MemTotal row) and an overhead row; enrolment comes after both.

[gunbai-bot]

At 9cb4551, one census receipt now yields both memory rows. census_meminfo_reading parses the capture's meminfo section once. host_usable_ram_from_capture takes MemTotal from it, and host_fixed_overhead_from_capture takes MemTotal − MemAvailable through host_overhead_upper_bound, the same bound srv3's row uses; a census boot is the quiescent read that bound's note asks for. The accept claim now reads overhead from the same designed receipt, so the designed allowance is gone. It still derives runner_count=4, and the new census_meminfo_receipt_reads_as_measured_fixed_overhead holds. All 10 claims evaluated true on BuildBuddy. Enrolment frontier: ONE trigger, the census boot's meminfo receipt. The RAM-backed tmpfs workspace charge is still open, and must land before enrolment.

[gunbai-bot]

At 73db6ac the tmpfs workspace is modeled in the slot envelope, no longer only noted in a comment.

• HostWidthFacts.workspace: HostWorkspaceStanding. Its production value, host_workspace_standing, reads the host's backing and, for a slot-charged (RamBackedTmpfs) backing, the workspace-inclusive peak through gunbc.runner_throughput_qualification aggregate_memory over the host's recorded qualification runs. No runs are recorded yet, so the value is MemoryUnread.
• committed_width_of refuses WidthWorkspacePeakUnestablished unless that peak is PeakWithinSlotHigh. Disk-backed hosts are unaffected: the probe still gives srv1=10 srv2=unresolved srv3=17 srv4=17.
• To avoid an import cycle: the mtcollins1_qualification_workspace_backing row (with its note) moves to gunbc.runner_throughput_qualification, and selected_slot and slot_quota_threads move to gunbc.runner_slot_allocation. Each still has a single definition.
• New RED mtcollins1_RED_tmpfs_host_with_no_qualification_peak_refuses: the same supplied facts with the production workspace standing refuse by type, in both width and placement. The accept claim now supplies a designed peak within the high and still derives runner_count=4.
• Frontier, in order: (1) census boot → RAM and overhead; (2) wet qualification receipt → workspace-inclusive peak.
  19 claims evaluated true on BuildBuddy at this head: the 11 offer claims, 3 host_width_facts claims, 2 slot_allocation claims, 2 qualification claims, and live-plan conservation. The rest of those suites are left to CI.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-09T17:39:41.447585Z	73db6ac	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 73db6acaae

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Match recorded runs to the full qualification subject

Once a run is recorded, this filter accepts every run from the host regardless of workload revision, runtime image, effective slot properties, or workspace locus. host_workspace_standing passes these runs directly to aggregate_memory without the subject_join enforced by assess_runner_throughput, so an old or NVMe-backed run with a low peak can yield PeakWithinSlotHigh and unblock committed_width_of for the RAM-backed tmpfs configuration even though that configuration was never qualified. Filter or assess runs against the complete selected subject before using their peaks for admission.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Include the new override in the declared-absence census

Adding this third SessionsOperatorDeclaredAbsent row leaves operator_declared_absent_host_count() filtering only srv1–srv4, so it still returns 2 and the existing witness remains green. After srv3 and srv4 gain observations, that census will report zero—and its scaffold can appear discharged—even while mtcollins1 still has an unobserved, capacity-granting override. Include mtcollins1 in that roster and update the expectation, or track this row under a separate typed dissolution condition.

Useful? React with 👍 / 👎.

[gunbai-bot]

Closed without folding in the v1 closeout bankruptcy (#13641). Waits on a census boot measurement that doesn't exist. Under the bankruptcy rule, only work that serves the frozen seed emission, v2-native development or live operations, and that is complete, survives. The branch is kept for archaeology; no follow-up obligation is created. — sent from neat-wolf-604