Ground fleet DIMM memory via cited HostMemoryPopulation (srv3 observed)

[gunbai-bot]

Summary

Dissolves the ComputeHost.memory hand-literal fork for srv3 by observing BMC Redfish per-DIMM inventory, citing DramModuleCatalogRow rows, and deriving MemoryDevice capacity through HostMemoryPopulation (module × stick_count).

• srv3 grounded (2026-07-16 BMC observation, 192.168.1.192 factory cred): MemorySummary 128 GiB; per-DIMM: 7× HMA82GR7AFR8N-VK + 1× HMA82GR7AFR4N-VK (mixed rank composition the old gibibyte(128) literal was hiding)
• New catalog row: hma82gr7afr8n_vk_catalog in dag/extdeps/memory/sk_hynix.dag (AFR8N = x8 2R: chip_width(8), rank_count(2))
• Existing row reused: hma82gr7afr4n_vk_catalog (AFR4N = x4 1R: chip_width(4), rank_count(1))
• ci_fleet.dag: gunbc_ci_host.memory now derives from srv3_memory_population (fixes wrong 125 GiB byte literal → cited 128 GiB)
• srv1/srv2: typed deferral rows (fleet_host_memory_population_deferrals) with dissolution trigger = rotated BMC creds for 192.168.1.183/.184; literals retained until observation lands as pure data rows

Redfish / part-number decode (no silent reconciliation)

Slot	PartNumber	Redfish RankCount	Redfish DataWidthBits	Catalog decode
dimm0 (×7)	HMA82GR7AFR8N-VK	2	64	x8 2R (chip_width(8), rank_count(2))
dimm7 (×1)	HMA82GR7AFR4N-VK	1	64	x4 1R (chip_width(4), rank_count(1)) — consistent with existing cited row

Earlier "1Rx8" wording for AFR4N was imprecise: JEDEC suffix AFR4N decodes to x4 single-rank, not x8 chip width.

Collateral

• std.machine_shape.dag: extract nested fold into helpers (unblocks gunbc run compile through storage→machine_shape import closure; pre-existing parser limitation on main)

Test plan

gunbc run --source-root dag --source-root src/v2 --claim-run \
  --entry dag/test/claim/fleet_intent_memory_witness_test.dag --function <each>

All 7 witnesses green:

• srv3_population_matches_bmc_memory_summary
• srv3_host_memory_derives_from_cited_population
• srv3_observed_catalog_rows_are_jedec_realizable
• srv3_population_row_bytes_match_stick_capacity
• red_wrong_stick_count_changes_nominal_total (7-stick-only ≠ 128 GiB observed)
• srv1_srv2_deferrals_are_counted_and_named
• srv3_observation_note_records_raw_redfish_fields

[gunbai-bot]

Addressing review feedback:

review 39028 (REQUEST_CHANGES) — fixed in 341aba3

Valid finding. gunbc_ci_host ("gunbc-ci-fleet") models the srv1/srv2 CI runner budget identity (roadmap: 128c/125 GiB; floor peak receipts from srv1/srv2 runners), not srv3 hardware. Wiring srv3_memory_population into ci_fleet.dag:123 was cross-host fact smuggling — reverted to byte_size(134217728000) (125 GiB). srv3 grounding remains on srv3_host only; gunbc_ci_host stays on the srv1/srv2 deferral path until those hosts are independently observed and cited.

review 39027 (APPROVE) — acknowledged, no code change

Both minor observations accepted as deferred:

• data_rate_mts: 2666 flat scalar mirrors the pre-existing hma82gr7afr4n_vk_catalog row; DRAM data-rate carrier is a catalog-schema tracker, out of scope here.
• deferral-row bmc_host IPs are ephemeral Scaffold rows with named dissolution triggers, not durable authority.

— sent from royal-tern-824

[gunbai-bot]

review 39033 (APPROVE) — acknowledged, no further code change from this review.

Verified against current HEAD (7b25c8e8f3): srv3 HostMemoryPopulation grounding, srv1/srv2 deferral rows, gunbc_ci_host kept on 125 GiB srv1/srv2 budget identity, and witness corpus all match the approval summary. Subsequent commit 7b25c8e8f3 (post-review) only extends deferral coverage per manager steer — names gunbc_ci_host.memory as fold-in site and records 125 GiB usable vs 128 GiB nominal as distinct facts; no scope change to the approved architecture.

— sent from royal-tern-824

[gunbai-bot]

review 39037 (APPROVE) — acknowledged, no code change.

Verified on HEAD 7b25c8e8f3: srv3 cited HostMemoryPopulation, srv1/srv2 deferrals with dissolution triggers, nominal-vs-usable distinction for gunbc_ci_host.memory, and machine_shape fold extraction all match the approval summary.

— sent from royal-tern-824

[gunbai-bot]

review 39041 (APPROVE) — acknowledged, no code change.

Verified on HEAD 7b25c8e8f3: srv3 cited population, srv1/srv2 + gunbc_ci_host.memory deferrals with nominal-vs-usable separation, witness corpus, and machine_shape fold dedup all match the approval summary.

— sent from royal-tern-824