feat(agents): persist agent capabilities

[casey-brooks]

Summary

• bump agynio/api buf dependency and regenerate bindings
• add agent capabilities column and store serialization
• map agent capabilities through create/get/list/update requests

Testing

• go test ./...
• go build ./...

Fixes #133

[casey-brooks]

Test & Lint Summary

Commands:

• go test ./...
• go build ./...

Tests: Passed: 1, Failed: 0, Skipped: 0.
Lint: not configured (not run).

[noa-lucent]

Clean implementation that follows existing codebase patterns well. Key observations:

• Migration: Correct — JSONB NOT NULL DEFAULT '[]' handles existing rows properly.
• Store layer: encode/decodeCapabilities are well-structured. The nil-to-empty normalization in encodeCapabilities correctly prevents writing SQL/JSON null to a NOT NULL column. The decodeCapabilities invariant checks catch DB-level corruption.
• Server layer: Proto3 repeated field presence detection via req.Capabilities != nil is the correct approach, and the explanatory comment is appreciated.
• Converter: Defensive append([]string(nil), ...) copies at the boundary are good practice.
• SQL: Column ordering, scan ordering, and parameter counts all verified correct.

One minor suggestion for unit test coverage on the encode/decode helpers. LGTM overall.

[noa-lucent]

[minor] decodeCapabilities and encodeCapabilities introduce JSON serialization logic that warrants unit tests — especially the edge cases:

• decodeCapabilities(nil) → error (NULL invariant)
• decodeCapabilities([]byte("null")) → error (JSON null invariant)
• decodeCapabilities([]byte("[]")) → empty []string{}
• decodeCapabilities([]byte(["a","b"])) → ["a","b"]
• encodeCapabilities(nil) → []byte("[]") (nil normalization)
• encodeCapabilities([]string{}) → []byte("[]")

These are easy to test and document the contract for future readers.