fix(responses): dispatch SSE on payload type, not just event: line (#434)

[initializ-mk]

Fixes #434.

Problem

The openai-responses SSE parser (readStream) routed events only on the SSE event: line. That field is optional per the SSE spec, and the Bedrock openai-sigv4 gateway (Kong /bedrock/openai-sigv4/v1/responses) omits it — it streams the event type as a type field inside each data: payload instead:

data: {"delta":"Loading","type":"response.output_text.delta"}
data: {"response":{...,"usage":{"input_tokens":76,"output_tokens":73,"total_tokens":149}},"type":"response.completed"}

With no event: lines, currentEvent stayed "", the switch matched nothing, and every frame was dropped — deltas and the terminal response.completed (usage).

Field evidence

• llm_call: input_tokens:0, output_tokens:0, tokens_unavailable:true on a 619ms call.
• Log: {"finish_reason":"","msg":"llm response"} — empty finish reason is the smoking gun (only response.completed sets one).
• User got an empty answer; the raw stream shows the gateway did return content + usage.

Fix

1. Dispatch on the payload type field, falling back to the event: line only when the payload has none (authoritative, present in both dialects; matches how the OpenAI SDKs route).
2. Bump bufio.Scanner's 64KB line cap — the response.completed frame embeds the full output[] on one data: line and overruns it on large answers.

Tests

• TestResponsesClient_DispatchesOnPayloadType_NoEventLines — the gateway dialect (no event: lines) recovers text + usage (76/73/149) + finish reason stop.
• TestResponsesClient_EventLineDialectStillWorks — real-OpenAI dual dialect still parses (no regression).
• All existing responses/provider tests pass; golangci-lint clean.

Out of scope

• The model emits read_skill("weather") as plain output_text in a phase:"commentary" message rather than a function_call item, so the skill never loads — a separate model/gateway-prompt behavior issue.
• Optional follow-up: capture usage.input_tokens_details.cached_tokens for cache-usage parity with Anthropic llm_call undercounts tokens: capture cache_read/creation + emit total_input_tokens (parity with initializ-sdk#10) #431.

https://claude.ai/code/session_01Hkimw1PDJRY5Dh8BgNQxWJ

[initializ-mk]

Approve-grade — precise diagnosis, minimal robust fix, no regression path. Traced the full readStream loop against the branch source.

Correct root-cause fix

The event: line is optional per the SSE spec, and the Bedrock openai-sigv4 gateway omits it, carrying the type only as an in-band type field — so the old dispatch left currentEvent=="" and dropped EVERY frame (empty answer, zero usage, empty finish reason, exactly the field symptoms). Preferring the payload type and falling back to the event: line is the right call: type is present in both dialects and is how the OpenAI SDKs route.

Things I verified beyond the happy path:

• No currentEvent-reset regression — eventType is recomputed per data: frame; since type is present in both dialects the fallback to a possibly-stale currentEvent effectively never fires, and where it would, the old code dropped the frame anyway.
• Non-JSON / [DONE] frames are safe — json.Unmarshal on a non-JSON data: returns an error (never panics), so eventType falls back to currentEvent and the switch no-ops. Correct terminator handling.
• Scanner buffer bump uses the right signature (Buffer(initial, max)) and fixes the real response.completed large-frame overrun.
• Regression coverage — the event-less gateway dialect recovers text + usage (76/73/149) + finish reason stop, and the real-OpenAI dual dialect still parses. Both green.

Non-blocking trivia

• The 8 MB line cap is a theoretical ceiling — a response.completed embedding >8 MB of output[] would still abort the stream. Generous and bounded, so fine in practice; awareness note only.
• The per-frame double-unmarshal (type probe + full payload) could be collapsed, but the lightweight probe-first approach is cleaner and the cost is negligible for SSE. Not worth changing.

The two deferred items (plain-text read_skill, cached_tokens #431 parity) are legitimately out of scope. All 9 CI checks green.

[initializ-mk]

This is the crux and it is correct. Preferring the in-band type over the event: line is what recovers the gateway dialect, and the guard is well-formed: on a non-JSON data: frame (e.g. a [DONE] sentinel) json.Unmarshal returns an error — not a panic — so eventType cleanly falls back to currentEvent, and && typed.Type != "" avoids clobbering a valid event:-derived type with an empty payload type. The result is robust across all three inputs: gateway (type only), real-OpenAI (both, matching), and junk/terminator frames (neither → no-op).