refactor(streaming): adopt the shared Praxis SSE codec

[leseb]

Description

After praxis-proxy/praxis#986 provides the shared bounded SSE codec, migrate
Praxis AI's provider-neutral SSE framing code to that praxis-filter API.

Praxis AI currently has several independently evolved framing paths:

• apis/src/openai/sse/frame.rs
• apis/src/openai/sse/responses/parser.rs
• apis/src/openai/responses/stream_events/
• apis/src/anthropic/stream_events/
• filters/src/agentic/a2a/sse.rs
• the A2A parser reuse in filters/src/token_usage/count.rs

The refactor should share only the transport-independent SSE record framing.
Provider event schemas, JSON decoding, lifecycle validation, accumulation,
translation, failure policy, response IDs, and sequence handling remain owned
by their existing AI modules.

Acceptance Criteria

• Update the Praxis dependency to a revision containing
  filter: add a bounded incremental SSE codec praxis#986.
• Replace apis/src/openai/sse/frame.rs with the shared incremental decoder and
  encoder, removing the local implementation once all callers are migrated.
• Migrate the Responses, Anthropic, A2A, and token-usage consumers where their
  framing requirements match the shared contract.
• Do not introduce a common provider event model; each consumer retains its
  typed event parsing and lifecycle rules above the shared SSE records.
• Preserve each consumer's existing pass-through, transformation, fail-open or
  fail-closed, terminal-event, and end-of-stream behavior.
• Preserve raw event bytes where a pass-through consumer requires byte-exact
  forwarding. If the shared codec cannot support a required behavior, refine
  filter: add a bounded incremental SSE codec praxis#986 rather than adding another AI-local framing parser.
• Retain consumer-specific limits such as event count, accumulated state, JSON
  size, and lifecycle timeouts; the shared framing bounds do not replace them.
• Tests cover arbitrary chunk splits, UTF-8 splits, LF/CRLF, comments,
  multiline data, multiple events per chunk, incomplete EOF, malformed input,
  and framing-limit overflow through the migrated production callers.
• No migrated streaming path buffers the complete response or weakens
  downstream backpressure.
• Remove obsolete duplicate parser tests and replace them with focused adapter
  tests plus the shared codec's conformance coverage.

Non-Goals

• Moving OpenAI ResponsesState, Anthropic translation, A2A task semantics, or
  token accounting into Praxis.
• Implementing GET /v1/responses/{id}?stream=true, event persistence,
  starting_after, or replay journals.
• Changing Pingora transport, BodyMode::Stream, or
  StreamingTerminalResponse.
• Combining provider-specific stream filters into one filter.

Dependencies

• filter: add a bounded incremental SSE codec praxis#986 — shared bounded incremental SSE codec

Parent Epic

• Epic: add shared Server-Sent Events stream primitives praxis#985

Related

• Filter 6: stream_events #25 — native OpenAI Responses stream_events
• SSE streaming inspection filter #89 — SSE streaming inspection
• Valid multi-line SSE data events are silently discarded #422 — multiline SSE data regression coverage