feat: type reasoning items and fail closed on corrupt stored history (Phrase 3 of #335)

[Zheng-Lu]

Part of #335 (split from #350)

Summary

• Replace the untyped JSON in ReasoningOutput with typed reasoning content, summaries and status, plus a bounded OpaqueReasoning string for encrypted_content: redacted Debug, 16 MiB ceiling. Public JSON field names are unchanged. The shared retained-response budget now charges reasoning parts and opaque bytes.
• Keep the existing vLLM replay policy: plaintext is joined, summaries are left out of the upstream copy, and opaque-only continuation is rejected before inference.
• Stop silently dropping stored context. Undecodable history rows, missing referenced rows, and malformed history references or response metadata now return typed storage errors. The errors never include stored values. SQL NULL keeps its legacy meaning.
• Keep reasoning rows written by earlier releases continuable. When a stored reasoning row fails typed decoding, a bounded projection keeps its plaintext (as reasoning_text) and every summary, string opaque state and status that still decodes. Only rows that cannot be read even this way fail closed.

Compatibility: request input and upstream output now reject reasoning items that don't match the typed schema, for example a non-string encrypted_content or an unknown status; earlier releases accepted them. No migration is needed, and existing rows are not rewritten.

Test Plan

• cargo test -p agentic-server-core -p agentic-server: 1,742 passed, 0 failed, 17 ignored. cargo clippy --workspace --all-targets -- -D warnings, cargo fmt --check and pre-commit (including Rust file sizes) pass.
• legacy_reasoning_rows_test.rs rewrites stored rows to shapes earlier releases accepted:
  • Each row decodes with its valid fields kept.
  • A previous_response_id continuation completes and replays the plaintext to vLLM as reasoning_text.
  • Negative control: with the projection disabled, both tests fail.
• reasoning_types_test.rs covers schema rejection, nullability, exact opaque round trips, redaction and the size ceiling. storage_response_integrity_test.rs and the storage suites cover invalid rows, missing rows, bad references and bad metadata, all with redacted errors.
• The existing recorder-generated OpenAI and gateway reasoning cassettes replay unchanged. No cassette was added or edited.