TAS-148: Shared-state concurrency challenge matrix and deterministic profile verdicts

[AtlantisPleb]

Summary

Public Tassadar needs shared-state concurrency challenge matrix and deterministic profile verdicts so the executor lane can widen honestly without collapsing profile boundaries, portability, or refusal posture.

Roadmap position

• Tranche: Tranche L | Semantic-window evolution and universality-governance pre-closeout
• Strategic phase: Phase 9/10 governance bridge
• Tranche goal: Govern semantic drift without losing honesty before the terminal contract.
• Why this tranche exists: This tranche exists so widening can continue without silent semantic or claim drift.
• Dependency context: Sequences after TAS-147 inside Tranche L and still depends on the landed baseline through TAS-102.

Problem

If concurrency or shared-state profiles ever widen, the system needs a challenge matrix that says exactly which scheduler and memory-ordering regimes remain honest.

Hypothesis

A concurrency challenge matrix with explicit deterministic verdicts will make shared-state widening either publishable under narrow envelopes or explicitly refuse broad claims.

Public source anchors

• docs/ROADMAP_TASSADAR.md
• docs/ARCHITECTURE.md

Target surfaces

• psionic-runtime
• psionic-provider
• psionic-eval
• psionic-research
• psionic-cluster

Claim class

• execution truth / refusal truth / operator truth

Initial scope

• implement the smallest honest subsystem changes across psionic-runtime, psionic-provider, psionic-eval, psionic-research, psionic-cluster that close the named gap
• add deterministic fixtures, receipts, reports, or capability rows that make the new posture machine-legible
• update public docs and capability language only where the new bounded surface is actually supported
• keep unsupported shapes on explicit typed refusal paths rather than widening implicitly

Validation and benchmark expectations

• shared-state challenge suites
• scheduler-order audits
• replay-versus-race reports
• and deterministic-profile verdict publication by concurrency class
• add negative coverage for unsupported or out-of-profile behavior
• update any machine-readable receipt, report, capability row, or acceptance artifact touched by this lane

Claim-discipline requirements

• keep the capability bounded to named profiles or declared semantic windows
• do not let this issue imply arbitrary Wasm, broad internal compute, or broader served posture unless an explicit promotion gate is part of the task
• preserve explicit refusal reasons for unsupported module shapes, effects, portability envelopes, or semantic families
• treat post-core proposal families, concurrency, and accelerator semantics as separate claim classes rather than inherited core-Wasm support

Done when

• the named surfaces land with deterministic tests, fixtures, or reports that exercise the new bounded capability
• public docs reflect the new capability, limit, or refusal posture in bounded language
• the claim remains benchmarked or promoted only at the level explicitly supported by the landed evidence

Status

• seeded