fix(claude): align Messages request and stream wire contract

[rexdotsh]

Summary

• Enforce Claude Code Messages headers and OAuth endpoint query; keep the existing broad OpenCode system-prompt workarounds for primary agents and subagents.
• Supply the Claude OAuth identity when no system prompt exists, round-trip colliding custom tool names while preserving server tools, and keep claude-code, OAuth, and interleaved-thinking betas in the default set. Fine-grained tool streaming is no longer added by default (caller-supplied betas still pass through).
• Preserve unknown and partial SSE data, bound event buffering, classify malformed/truncated streams and mid-stream errors, flush partial usage on cancellation/failure, and record stream terminal outcomes. An incomplete EOF is logged and forwarded without a fabricated error event so OpenCode can classify it as a retryable interrupted stream. Incrementally track event bytes and scan only newly appended SSE data to avoid rescanning/re-encoding a growing event on every chunk.
• Stop writing unused provider-account metadata: Claude wire-config fields, Codex ID token/organization IDs, request-profile placeholders, and token type/scope. Legacy JSON remains readable and is sanitized in application memory. Existing stored rows are left for a separate manual cleanup later. Routing/refresh-critical fields remain.

Rebased onto main after Copilot removal in #32; Copilot accounts and usage rows were cleaned up separately after deployment.

Beta comparison: Pi removed fine-grained tool streaming on June 8, 2026 after large tool inputs produced malformed or wrong-shape JSON. Griffin’s current base betas and the community OAuth plugin also do not add it by default. OpenCode V2 maps @ai-sdk/anthropic to the native Anthropic provider, whose Messages transport defaults to interleaved thinking rather than fine-grained tool streaming. Kleis currently adds fine-grained streaming itself; this PR stops adding it by default. Explicit caller-supplied betas still pass through.

Supersedes #48; related to #37.

Validation

• bun test (158 passed after rebasing on perf(proxy): read and parse request bodies once #51)
• bun typecheck
• bun lint
• bun run build
• git diff --check
• Synthetic 1 MiB SSE event transformation, 1 KiB chunks: ~140 ms before incremental buffering vs ~55 ms after; 8 KiB chunks: ~21 ms vs ~10 ms (three local runs each). Not an end-to-end latency measurement.
• Compared with current main in the same local synthetic harness (after warmup): 1 MiB fragmented event at 1 KiB chunks ~72 ms main vs ~52 ms PR, at 8 KiB chunks ~14 ms vs ~9 ms. A 1,000-event small-text stream was also faster in these runs. These are CPU microbenchmarks, not model/network latency.
• Live local dev-server + OpenCode v2.0.14 with Opus 5: text reply and read-tool round trip successful before rebase and incremental buffering.