Refuse unresolved and uncited Altra 2DPC operating rates

[gunbai-bot]

An empty rank table must not turn the cited DDR4-3200 platform ceiling into an operating rate for every two-DIMM population. The uncited 2933/2400 transcription remains retracted. Missing rank-specific rows now produce a typed unresolved clock and bandwidth, carrying DIMMs per channel, rank count, and the platform maximum.

Matching rows must also carry CitedToAuthority evidence. TranscribedUncited rows refuse with their read obligation; missing rows and uncited rows have distinct causes. The admitted-path fixture cites its own synthetic data and establishes no Ampere operating rate. Gated AVL/layout material remains unread and proves neither a derate nor its absence.

The controls cover empty tables, nonempty uncited tables, and matching cited fixture rows capped by the platform maximum. #11068 owns the separate same-channel 1R/2R population constraint: this change does not alter that rule or turn different-channel TBD into a speed result.

Mutation validation at abad65677336034460dd2bcae7f43a46111223b0 was run by parent session crisp-lynx-452 in an isolated detached worktree using /cargo-target/release/gunbc. Its ProcessExit driver checks emptying_rank_specific_ceiling_rows_does_not_admit_an_effective_rate (ARM1) before uncited_rank_specific_rows_do_not_admit_an_effective_rate (ARM2), with named failure reasons. The production file was restored from a pristine copy between mutations. Run receipt: dashboard message msg_44d1f664-cf18-4212-bd28-5f069a17fa67 (logs retained by that session).

Run	Exit	Result
Intact implementation	0	Both controls pass
Restore empty-table fallback: Absent => ChannelClockAdmitted { rate: platform_max }	1	ARM1 EMPTY_ROWS fails
Ignore provenance: TranscribedUncited => ChannelClockAdmitted { rate: min_rate(a: ceiling, b: platform_max) }	1	ARM1 passes; ARM2 UNCITED_ROWS fails

Thus each control passes intact and detects its corresponding faulty arm. The provenance mutation also demonstrates that the empty-table control alone cannot catch uncited-row admission. The uncited control asserts RankSpecificRateUncited and its read obligation, so refusal with the wrong cause does not pass.

Earlier infrastructure attempts in smart-crab-414 did not evaluate controls: the remote compiler built, but the interpreter exited 101 with HostBudgetUnreadable; the local build exited 101 because this container cannot write /cargo-target/release. These remain infrastructure findings, not semantic reds.

Required CI on this exact head and the operator's re-review remain pending. The operator's CHANGES_REQUESTED is not cleared by a push.

Static validation: git diff --check passed. No effective speed is established for the fleet purchase by this PR.

[gunbai-bot]

review 63200 is right: the one-constructor HeldDatasheetTwoDpcRankDerate and MtJadeSameChannelRankMix rows were only matched in witnesses, so those matches were permanently green.

On f88b3d0a9dd:

• Those types are gone. Datasheet absence is the empty ceiling table, which ceiling_for_ranks already consumes.
• Mt. Jade Table 3 is mt_jade_two_slot_rank_standing (uniform vs unlike ranks), consumed by channel_effective_data_rate / memory_bandwidth. 2DPC 1R+2R on one channel is TwoDpcUnlikeRanksOnOneChannelRefused — that RED is authorable (two_dpc_unlike_ranks_on_one_channel_are_refused).
• Held-unit mix is not an Mt. Jade fact; different-channel mix remains TBD and is not this function.

— sent from proud-carp-600

[gunbai-bot]

review 63205 — both findings held against the tree.

1. Deleted effective_data_rate. rate_at now matches channel_effective_data_rate (equal slot ranks). Production memory_bandwidth already routed through that function; it is the single clock.

2. Removed MemoryBandwidthRefusedUnlikeRanksOnChannel / MemoryBandwidthReading. PopulationOption still cannot express unlike ranks, so a bandwidth-level refuse would be a permanently green decoration. The authorable RED remains channel_effective_data_rate(..., slot0_ranks: 1, slot1_ranks: 2) plus mt_jade_two_slot_rank_standing. memory_bandwidth returns Bandwidth; the ChannelDataRate refuse arm is exhaustiveness only and reapplies the uniform-rank admitted clock for this one-module caller.

— sent from proud-carp-600

[gunbai-bot]

review 63209 — the fabricate arm was a real §5 widen; it is gone.

memory_bandwidth no longer matches TwoDpcUnlikeRanksOnOneChannelRefused. PopulationOption still carries one module, so a bandwidth-level refuse would be the permanently-green decoration review 63205 already flagged (equal ranks in, refuse arm uninhabited). Widening that type to two slot ranks is a provisioning-shape change, not this datasheet grounding.

The mix refusal stays where a mix can be authored: channel_effective_data_rate / mt_jade_two_slot_rank_standing, with the existing RED at two_dpc_unlike_ranks_on_one_channel_are_refused. The one-module clock is cited_channel_clock, consumed by both that function (after uniform standing) and memory_bandwidth.

— sent from proud-carp-600

[gunbai-bot]

review 63216 — both findings held.

• Deleted unused altra_memory_max_data_rate_citation (and the FactCitation import). The rate row stays; standing for the empty 2DPC table remains altra_two_dpc_ceilings_evidence.
• Renamed downclock_is_rank_dependent to two_dpc_keeps_the_cited_rated_rate_across_ranks so the name matches the 3200-across-ranks claim.

— sent from proud-carp-600

[gunbai-bot]

review 63225 — both findings held.

• altra_cited_rated_rate is now min(context.max_data_rate, altra_memory_max_data_rate). mt_jade_dimm_speed_at_two_dpc is no longer a term in that fold (it was also applied on 1DPC and on ALTRAD8UD). It remains a Mt. Jade row, asserted as corroboration in the empty-table witness (and mt_jade_platform_witness).
• Deleted rate_at -> Nat with 0 - 1. Uniform-rank checks use admitted_rate_equals -> Bool (DataRateAdmitted vs false on refuse). Mix RED still matches ChannelDataRate directly.

— sent from proud-carp-600

[gunbai-bot]

review 63231 — both findings held; did not widen PopulationOption.

• Deleted channel_effective_data_rate / ChannelDataRate. That wrapper had no production consumer (memory_bandwidth uses cited_channel_clock). Widening PopulationOption to per-slot ranks is a provisioning-shape change, not this datasheet grounding. Mt. Jade Table 3 stays in extdeps.ocp.mt_jade.platform mt_jade_two_slot_rank_standing, executed by mt_jade_platform_witness.
• CitedFigureStanding now lives on TwoDpcRateCeilingTable.evidence. Removed altra_two_dpc_ceilings_evidence and the unused per-row TwoDpcRateCeiling.evidence.

— sent from proud-carp-600

[gunbai-bot]

review 63273 — the finding held. Did not route Table 3 into memory_bandwidth.

PopulationOption still carries one module.rank_count. Wiring mt_jade_two_slot_rank_standing there would make SameChannelUnlikeRanksNotSupported uninhabited (review 63205 / 63209). Widening the population shape is a provisioning change, not this datasheet grounding.

Deleted TwoSlotRankStanding / mt_jade_two_slot_rank_standing and the witness-only matches. mt_jade_dimm_speed_at_two_dpc remains (3200 at 2DPC, corroboration of the empty Ampere ceiling table).

— sent from proud-carp-600

[gunbai-bot]

Mechanical merge of origin/main, pushed by session crisp-lynx-452 — no content change.

This PR's own lane (proud-carp-600) was idle through a delivered high-priority request and this PR
is slot 1 of a staggered push order, so quick-ibex-468 ruled that the merge be performed on its
behalf. Recording it here so the approvals on the prior head remain traceable.

What was done: origin/main (267d69b86cb) merged into session/proud-carp-600 as a merge
commit. No rebase, no force-push, no edit to any file in the diff.

b04a742e743 → 6d3fba0c5c5

Why the prior reviews still describe this head. The merge is verifiably content-free:

git diff origin/main...HEAD before merge	22,759 bytes — 7 files, 127 insertions, 90 deletions
git diff origin/main...HEAD after merge	22,759 bytes — byte-identical
merge conflicts	none

The subject the reviewers approved is unchanged. Had the merge produced any conflict, the instruction
was to stop and report rather than resolve it, on the grounds that a conflict would mean the change
was not content-free.

Freshness, by ancestry rather than push time:

gh api repos/gunb-ai/gunbc/compare/267d69b86cb...6d3fba0c5c5 --jq .status
  ahead   (ahead_by=10, behind_by=0)

The push was necessary because this head predated gunbc#10992, so its runs were reproducing an
inherited D0-ADJUDICATE refusal — four claims interrupted before verdict at 503–508 ms against a
500 ms CPU budget, claims_failed=0. A re-run could not have cleared it: the merge ref is pinned at
push time, so a re-run replays the stale ref.

Push-time timestamps were explicitly retired as a freshness test earlier tonight, after they
misclassified two stale PRs as fresh — a head pushed after a fix but merging an older main is still
stale. Ancestry is the test.

The PR remains proud-carp-600's. Nothing further will be done to it by this session.

[gunbai-bot]

review 63496 — held. Deleted altra_cited_rated_rate. cited_channel_clock uses context.max_data_rate only. The Altra 3200 citation stays on the context construction sites (altrad8ud_context / mt_collins_context).

— sent from proud-carp-600

[gunbai-bot]

review 63528 — held. altra_two_dpc_ceilings.evidence is now CitedToAuthority to mt_jade_subject_authority_anchor (rev 1.0 §7.1 "DDR4-3200 (2DPC)"), not the Altra datasheet. The datasheet remains silent on a rank-derate table; AVL/PCB stay unread and are not a licence to restore 2933/2400. The 3200 2DPC clock still comes from context.max_data_rate at construction, not a min of silence.

— sent from proud-carp-600

[briansrls]

HOLD: removing the uncited 2933/2400 table is correct, but this head replaces one unsupported inference with another.

“Up to DDR4-3200,” “up to 2 DPC,” and Mt Jade's feature-list phrase “DDR4-3200 (2DPC)” establish a platform capability/ceiling. They do not establish that every 1R, 2R, and 4R two-DIMM population operates at 3200. The absence of a published rank-dependent table is an unresolved input, not a cited empty table whose default is the maximum.

Today altra_two_dpc_ceilings.rows = [] plus cited_channel_clock makes the absence executable as 3200, and two_dpc_keeps_the_cited_rated_rate_across_ranks positively asserts 1R/2R/4R at 3200. That is exactly the inference the source does not carry. The gated AVL/layout material being unread cannot prove a derate, but it also cannot prove there is none.

Please:

• keep 2933/2400 retracted as uncited;
• represent the per-population effective rate as unresolved unless an exact authority or an identity-bound observation supplies it;
• keep the cited 3200 value as a maximum/platform capability, not a default row for every rank count;
• add a control showing that deleting all rank-specific rows does not manufacture an admitted effective rate;
• keep #11068's same-channel 1R/2R refusal as a separate population-admission constraint and do not turn different-channel TBD into a speed result.

This head is also behind current main and its required witness run is red; neither condition should gate #11066/#11068/#11071 or the unit-1 receipt recut.

[gunbai-bot]

Answering the operator CHANGES_REQUESTED on this PR (GitHub review at 17:38Z): empty rank-specific 2DPC rows no longer execute as DDR4-3200.

cited_channel_clock / memory_bandwidth now return Unresolved for 2DPC unless a matching rank row applies. The 3200 figures stay on the platform-capability carriers (altra_memory_max_data_rate, mt_jade_dimm_speed_at_two_dpc). The 2933/2400 transcription stays retracted. Control emptying_rank_specific_ceiling_rows_does_not_admit_an_effective_rate is green on this head (15/15 node_power_envelope witnesses).

— sent from proud-carp-600

[gunbai-bot]

review 63993 held on both findings.

• Deleted derived_bandwidth_count. memory_bandwidth_is_derived_from_the_populated_topology and a_single_rank_population_does_not_reach_the_rated_bus now match MemoryBandwidthDerived / MemoryBandwidthUnresolved; a refusal no longer greens the < comparison via 0 - 1.
• Enrolled witness-local fixture_admitted_two_dpc_ceilings (not Ampere, not the retracted 2933/2400). a_rank_specific_ceiling_row_admits_the_capped_effective_rate executes 1R unresolved, 2R admitted at min(4000, 3200)=3200, 4R admitted at 2500.

16/16 node_power_envelope witnesses green locally.

— sent from proud-carp-600