Roadmap: consequence standing, extension law, scenario evaluation, customer CI failover

[gunbai-bot]

Adds four roadmap nodes under fabric-bringup. They extend fabric-demand-objective (#13801) with what the operator side-chat discussion established and no PR yet carried.

Node	Depends on	Target
fabric-consequence-standing	fabric-demand-objective	Every candidate carries one disposition per required consequence dimension: observed, estimated, bounded, irrelevant or unestablished. Unknown setup or reconstruction is never zero.
fabric-dimension-extension-law	fabric-consequence-standing	A new dimension enters through the shared evaluation, never as a side controller. Every extension ships a must-change and a must-preserve control through the real decision path.
fabric-scenario-evaluation	fabric-consequence-standing, fabric-dimension-extension-law	Seeded joint scenarios (coupled facts drawn together) produce per-dimension outcome distributions. Old and new MODEL revisions are compared on identical seeds under a fixed decision authority, gated on safety, dominance, the extension law's preservation control and regret. The REDs are structural: an undeclared loss ratchet crosses an invariant, and ownership history changes a selection among equivalent offers.
fabric-customer-ci-failover	fabric-demand-objective, fabric-consequence-standing	Customer CI failover, balancing, cloud burst and return-to-owned-capacity are one allocation on the shared decision interface. The transition plan is part of each candidate; disruption stays decomposed; scarce capacity is selected under the affected demands' declared objectives and contracts, with normalized delay as an axis reading or constraint, not a universal rule. The red control is the eight-case receipt.

Each node states boundary, displaced cost, first slice, RED control, out of scope and handback. Customer CI failover joins the conformance-decision consumer list in that node's first slice; it is not added here, because #13825 is open against the same rows.

Verification: a remote __no_such_function__ run returns NoSuchFunction, so the roadmap authority resolves and typechecks. ROADMAP.md is not regenerated: #13810 (open) stops committing it, and the cutover owner asked lanes not to regenerate it meanwhile.

🤖 Generated with Claude Code

[briansrls]

Requesting bounded structural repairs at 4b49ac7.

The stack's top-down decomposition is right. #13801 is the sound root: the demand's declared objective governs resource choices. #13825 adds the shared onboarding/conformance interface. This PR then correctly separates (a) consequence standing, (b) the extension law, (c) deterministic joint-scenario evaluation, and (d) customer CI failover as another consumer rather than another scheduler.

Three issues block this head:

1. The roadmap dependencies do not close over what the nodes consume. fabric-customer-ci-failover requires the consequence-standing carrier in its own first slice (spend, time to accepted result, repeated work, liability, plus startup/storage unknowns), but it is held only by fabric-demand-objective. That permits it to dispatch before the shared consequence carrier and re-mint those dispositions locally. Add a dependency on fabric-consequence-standing. Likewise, fabric-scenario-evaluation claims the extension law's must-preserve gate; either hold it behind fabric-dimension-extension-law or state a single ownership direction so these do not become two preservation authorities.

2. Customer-CI arbitration hard-codes a local priority rule: Scarce recovery capacity ... goes by delay relative to each customer's own allowance. Delay relative to allowance is a useful axis, and a contract may explicitly select a max-normalized-disruption policy, but it is not the universal winner across customer objectives. As written this repeats the exact problem the shared interface removes: one local score stands above cost, accepted work, SLO constraints, release criticality and other contract terms. Carry normalized delay as a declared consequence/constraint; let the owning contract/objective select among the non-dominated allocation candidates.

3. Scenario regression needs a fixed decision authority. A new policy must beat the old one / regret no worse is only meaningful when subject, objective, constraints, candidate field and evaluation metric are held fixed. #13801 deliberately requires different objectives to produce different winners; a latency-sensitive policy must not be gated as a regression against debt-payoff under one shared regret score. Scope the paired gate to model/implementation revisions under the same decision authority, and treat a changed objective as a new decision. The historical controls should name discriminating scenarios/invariants rather than claim universal economic inferiority. In particular, history-keyed-limit ... loses to the current policy is presently ambiguous because the live ctrl policy introduced by #2333 is itself history-keyed.

Stack note: #13825 remains blocked on the one wording repair already requested there. A declared bounded feedback/update law must be allowed to consume a loss observation without an operator policy revision on every step; otherwise this PR's black-box/convergence work is classified as a permanent departure. #13801 is merged and remains the correct foundation.

No other blocker found. The consequence disposition, unknown-never-zero rule, joint-scenario requirement, transition-as-candidate framing, storage-as-supply treatment, fencing/readback cases, and deterministic seeded corpus all fit the intended system.

At review time witnesses and heal were still in progress.

[briansrls]

APPROVE at 542a07b.

The three requested repairs are present and close the prior blockers:

1. Dependency closure: customer CI failover now depends on consequence standing; scenario evaluation now also depends on the dimension-extension law, and the node explicitly states that the scenario runner is the implementation consumed by that law's preservation control.
2. Customer-CI arbitration: normalized delay is now an axis reading or contract constraint, not a universal score. Scarce capacity is selected over the complete candidate field under the affected demands' declared objectives and constraints; a contract may explicitly choose minimax normalized disruption.
3. Fixed-authority evaluation: paired scenario comparisons now hold subject, objective/valuations, constraints, candidate-field authority, horizon and outcome interpretation fixed. A changed objective is a different decision, and the historical REDs are structural rather than unsupported universal profitability claims.

I rechecked the complete one-file diff and edge additions. The four-node decomposition remains coherent: demand objective -> consequence standing -> extension law / deterministic joint-scenario evaluation, with customer CI failover consuming the shared objective and consequence carrier. No further blocker found.

Land once required checks pass; heal and witnesses were still in progress at my last check.