[TASK][MVP] Implement Provider Capability Assessment

[Noam-Alum]

Required result

Outcome and ownership

Implement locked P2 Capability Assessment as a private Provider runtime service. Authors declare dynamic prerequisites with @requirement_check(...) and reference stable IDs from @operation(..., requires=(...)); they never implement the SFP procedure or construct protocol findings. Preserve P1 manifest schema, R2 readiness, and the P2 wire format.

Provider author API

• Add public @requirement_check(identifier, description, version_constraint="") for bounded async, read-only checks.
• Add requires=(...) to @operation. Keep the registry reusable by future R4 trigger declarations, but do not implement triggers here.
• Expose typed author results: satisfied, unsatisfied, incompatible, and unknown. Do not expose P2 messages, findings, sessions, correlation identity, or envelopes.
• Discover checks through the existing declaration model. Publish static metadata through existing manifest runtime_requirements; keep operation associations Provider-owned for MVP and do not revise manifest schema 1.
• Fail startup for non-canonical/duplicate IDs, duplicate/unresolved references, invalid async signatures, or metadata outside Protocol bounds.

Private service

• Implement _services/capability_assessment on the multiplexed Provider endpoint.
• Delegate context-independent request/response validation to canonical Protocol validation. Provider owns only active session/generation, served manifest revision, declared operations, readiness, check execution, and adaptation.
• Treat execution_context as bounded opaque metadata. P2 has no generic target, operation input, or permission semantics; invent none and perform no target mutation.
• Collect requirements for requested operations, deduplicate by ID, and execute each distinct check at most once under RPC deadline/cancellation. Checks never dispatch operations or mutate targets. Reuse results across operations, preserve request-correlated ordering, and revalidate session/readiness before return.
• Aggregate: all satisfied → SUPPORTED; unsatisfied → SUPPORTED_WITH_REQUIREMENTS; incompatible → UNSUPPORTED; unknown → INDETERMINATE unless incompatibility already proves UNSUPPORTED. Negative/unknown results are domain outcomes. Invalid returns or unhandled check exceptions produce sanitized INTERNAL, never silent UNKNOWN.
• Preserve declaration vs readiness vs dynamic prerequisite vs point-in-time assessment vs later authorization/admission.

Documentation

• Replace obsolete assess_capability guidance in .claude/PROVIDER_AUTHORING_VISION.md.
• Document @requirement_check, operation requires, result meanings, startup validation, deduplication, read-only/error semantics, internal service ownership, and static manifest metadata.
• Include a complete filesystem-writability or service-control example. Do not edit or regenerate the PDF.

Verification

• Test declaration validation, dedup/result reuse, aggregation, ordering, deadlines/cancellation, invalid returns, sanitized exceptions, and no dispatch/mutation.
• Test Protocol validation, limits, stale session/generation, manifest mismatch, undeclared operations, readiness loss, response validation, and error mapping.
• Run a focused real Provider service scenario over a real Unix socket with exact Provider/Protocol artifacts. Preserve installed-Agent regressions, but A3 owns real Agent P2 consumption.
• Pass formatting, lint, strict typing, docs, unit/integration, packaging, and applicable CI.

Non-goals

No wire redesign, manifest revision, public assess_capability hook, author-built findings, target model, authorization/admission, triggers, operation execution, SF-AOP, Agent P2 client, or PDF update. Preserve repository ownership and human delivery authority.

Implementation step

R3

Closes #23