Implement the real OpenCode event contract; reject headless questions (#28)

[TheRealDinghyDog]

Fixes #28 (review findings H-03 and M-02).

Problem

The SSE client parsed an invented event shape. Verified against the real 1.17.15 OpenAPI document:

• message.updated carries only message metadata in properties.info — the client looked for parts arrays on it that never exist.
• Streamed content arrives exclusively via message.part.updated (properties.part) and message.part.delta, which the client never handled — so no live text, reasoning, or structured-output part was ever captured from the stream, and a valid answer was reported as failed whenever the held-open /message response dropped and HTTP recovery was unavailable (the review's targeted reproduction).
• turnId could pick up the evt_… envelope id instead of the assistant message id.
• Child sessions announce themselves via session.created with properties.info.parentID, which the client didn't read.
• question.asked/question.v2.asked had no handler: a plan-agent question stalled the turn for the full 30-minute ceiling (M-02, reproduced in the review).

The fake test server emitted the same invented shapes, so the suite reinforced the bug.

Fix

Client (opencode.mjs, opencode-server.mjs):

• Part snapshots accumulate by part id (insertion order preserved); the main-session final message is assembled from the turn's assistant-message text parts. Deltas append to known parts; the next snapshot stays authoritative — so mixed snapshot/delta streams self-correct.
• message.updated is metadata-only: assistant message id (fixes turnId) and info.structured for json_schema output.
• session.created/session.updated register children from properties.info; subagent text is labeled in progress logs and excluded from the final message.
• Questions are rejected immediately via the new rejectQuestion() (POST /question/{requestID}/reject, no body — per the OpenAPI doc) with a diagnostic naming the question; deterministic headless policy per the review's recommendation.

Contract pinning (the review's "generate fixtures from a captured /doc schema"):

• tests/opencode-event-contract.json is pinned from a real 1.17.15 /doc: envelope requirements, per-event properties required/allowed keys, and required keys for Session/AssistantMessage/UserMessage/TextPart/ReasoningPart/ToolPart.
• The fixture now emits the real envelope ({id: "evt_…", type, properties}) with contract-complete objects, and a new contract test drives it over raw HTTP, validating every emitted event against the pin — the fixture can no longer drift back to invented shapes.

Tests

• New: question rejection (turn completes, rejection recorded — previously a 30-minute stall); delta-only final-message assembly with the POST dropped and recovery withheld (the exact pre-Event client does not implement the current OpenCode message contract; question events stall headless turns #28 data-loss scenario); subagent output isolation; full fixture-vs-contract validation.
• All issue Transport resilience: spurious fetch failed on long turns + no result recovery #2/Transport: /message POST held open for the whole turn — long turns fail to return (follow-up to #2) #12/Transport hardening (post-#12): SSE-drop loses a valid turn; recovery has no timeout #15 transport fail-mode regressions re-verified under the real shapes.
• npm test: 89 passed, 0 failed, 0 skipped.

Not covered here: version-matrix compatibility testing against a minimum supported OpenCode (tracked as part of #33's CI work); a provider-backed live turn remains on the release checklist.

🤖 Generated with Claude Code