Floor: upload the cross-claim demand census from the required floor job

[gunbai-bot]

Part (b) of FLOOR REPAIR C3 (plan #13030): gunbc.floor_cost_distribution_instrument lost its inputs when the required floor moved into the job emitted by gunbc.compiler_gate_workflow, because the TSV uploads were authored only in gunbc.witness_floor_workflow's step list.

What this change is now

Main (#13227) has since restored the per-claim cost receipt upload through one bound function, required_floor_claim_cost_upload_bound_step. This PR is narrowed to what is still missing: the cross-claim demand census upload.

• required_floor_cross_claim_demand_upload_bound_step in gunbc.witness_floor_workflow, the same shape as CI: restore the required-floor-claim-cost upload; RFM row for an instrument artifact that stopped existing while cited #13227's function.
• Bound by name in both floor jobs: the witness floor's step list (replacing its inline copy) and compiler_gate_floor_measurement_bound_steps.
• .github/workflows/witnesses.yml gains the one step, after the cost receipt upload. The job already sets GUNBC_REQUIRED_FLOOR_CROSS_CLAIM_DEMAND, so the floor already writes the file; nothing collected it.

The earlier form of this PR bound all five TSV uploads through one list. That is dropped in favour of main's one-function-per-artifact pattern, so there is one way to bind a floor artifact.

Evidence

• The artifact exists on this head. Run 37201748835 (15036bd, all four jobs green, generated included, so the hand-added workflow step matches the .dag authority) uploaded required-floor-cross-claim-demand: 22,764 producer rows plus the summary line, 2.98 MB. The same run uploaded required-floor-claim-cost (671 claim rows) through main's CI: restore the required-floor-claim-cost upload; RFM row for an instrument artifact that stopped existing while cited #13227 step.
• Who consumes it, stated plainly. gunbc.floor_cost_distribution_instrument reads ONLY required_floor_claim_cost.tsv (its recipe and floor_cost_artifact_file say so); that input is restored by CI: restore the required-floor-claim-cost upload; RFM row for an instrument artifact that stopped existing while cited #13227, and the instrument was shown reading it on the earlier head (run 37078324678's artifact, 664 rows). No .dag instrument reads the cross-claim demand census. Its consumers are the recipes that cite it: docs/plans/witness-cost-first-touch-attribution.md, docs/plans/floor-cost-distribution-near-the-ceiling.md, and the attribution work in plan Floor time attribution: ranked structural causes (plan only) #13030, which is how the derived-share change (Floor: derive cross-claim pure-share from planned claims' call-site demand; delete the hand roster #13043) was sized. So this restores an artifact that stopped existing while cited (gunbc.recurring_failure_mode instrument_artifact_stopped_existing_while_cited), not an instrument input.
• Cost: about 3 MB per floor run at 14-day retention.

🤖 Generated with Claude Code

[gunbai-bot]

floor_cost_distribution_instrument is shown reading this PR's uploaded artifact.

This PR's run 37078324678 uploaded all five floor TSVs. The artifact list is required-floor-claim-cost, required-floor-cross-claim-demand, required-floor-disposition, expected-red-roster-join and long-home-storage-agreement, beside the receipt. The generated lane's drift gate passed on the hand-matched YAML.

The instrument was run on that artifact as follows:

• gh run download 37078324678 -n required-floor-claim-cost -D /tmp/floor-cost/37078324678
• tools.floor_cost_distribution_instrument floor_cost_distribution_report_for(root: "/tmp/floor-cost", run_ids: ["37078324678"], baseline: "37078324678", contended: "37078324678", repaired: [], order: [])
• It was evaluated with gunbc run from a scratch entry. The source is unchanged.

What it established:

• runs loaded 1, runs refused as observed 0, rows excluded ... right-censored 0.
• The band histogram covers every executed row: 540 + 77 + 14 + 22 + 11 = 664. That equals the artifact's # summary executed=664, so the current-vintage header (cost_reading plus the per-kind columns) decoded with nothing refused.
• The inflation percentiles here are self-paired (baseline = contended = one run), so they carry no information. Only consumption is being shown.

Not in this PR: the instrument's zero-argument entry still names ten runs that predate the per-kind columns, and its own header says they refuse. Re-pointing floor_cost_sampled_runs at current-vintage runs is a data edit for whoever re-runs the near-ceiling analysis. It needs a deliberate run pair, not this PR's run.

[gunbai-bot]

Closing, by sharp-raven-357's ruling (2026-10-04).

Part (b) of the brief, giving gunbc.floor_cost_distribution_instrument its input back, is met by main's #13227: that instrument reads only required_floor_claim_cost.tsv, and #13227 restored that upload through required_floor_claim_cost_upload_bound_step.

What this PR had narrowed to, the cross-claim demand census upload, has no executing consumer: its only readers are recipes in docs/plans/ and a merged plan. That is the DESIGN §3c dangling case, and §6 says an instrument is a named entry point, not an artifact scraped by hand. If a .dag instrument later needs the census, it adds the upload together with its reader in one PR; required_floor_cross_claim_demand_upload_bound_step on this branch is the shape to reuse.

— sent from royal-deer-478