…ix comment lines

- Replace ngx_re.split (PCRE) in sse.decode() with a pure Lua forward scan.
  Avoids regex compilation and PCRE overhead on every streaming chunk.
- Add sse.decode_buf(buf) that combines split_buf + decode into a single
  O(n) forward pass, halving the number of scans per chunk in hot paths.
- Fix SSE comment lines (lines starting with ':') being incorrectly treated
  as fields. Per the SSE spec, comment lines must be silently ignored; the
  previous code set has_field=true for them, producing spurious empty events.
- Add test: SSE comment lines are correctly discarded by decode_buf.

Co-authored-by: Copilot <[email protected]>

…path

- base.lua: replace sse_buf string concatenation with sse_parts table to
  avoid O(n) string allocation on every chunk; use table.insert + concat
  only when needed. Also adds ngx.sleep(0) yield at end of each iteration
  so health checks and concurrent requests are not starved under bursty
  SSE upstreams.

- base.lua: add streaming_flush_interval_ms support - when > 0, a
  background ngx.thread flushes the output buffer every N milliseconds
  instead of per-chunk. Useful when the upstream bursts multiple tokens
  at once and you want to bound client latency without synchronous flush
  overhead on every chunk. Thread spawn failure falls back to sync flush.
  flush_thread is killed on all exit paths (EOF, error, 502, limits).

- plugin.lua: add no_flush parameter to lua_response_filter(). When true,
  ngx.flush is skipped so the background interval thread can own flushing.
  Backward compatible - existing callers passing (ctx, headers, body, wait)
  are unaffected.

- schema.lua: add streaming_flush_interval_ms field (integer, min=0,
  default=0) to both ai_proxy_schema and ai_proxy_multi_schema.

- t/plugin/ai-proxy-flush.t: E2E tests for flush interval behavior,
  using error_log / no_error_log to assert whether the background flush
  thread fires. Also covers schema validation of negative values.

…blank-line chunks

decode() previously broke callers that pass full response bodies or
per-chunk bodies without a terminal blank line. The old ngx.re-based
decoder treated the remaining content as a complete event; the new
implementation discarded it.

Fix: when no boundary is found for the remaining content, parse it as a
final event via parse_raw_event instead of silently dropping it.

- Check ngx.flush(false) return in flush_thread; propagate error via
  shared flush_err upvalue so the main loop can abort cleanly
- Check flush_err at the top of the streaming loop and call
  abort_on_disconnect to stop upstream reads on client disconnect
- Skip final ngx.flush(true) when flush_err is set to suppress noisy
  'final flush failed' log on client disconnect
- Change streaming_flush_interval_ms default from 0 to 10 ms
- Add TEST 3/4 to ai-proxy-client-disconnect.t covering async flush
  disconnect scenario
- Minor ai-proxy-flush.t cleanups (explicit ms=0 in TEST 1, TEST 6 title)

Co-authored-by: Copilot <[email protected]>

… default

- Update base.lua comment: clarify default is 10ms, not 0; describe both
  modes (0 = per-chunk ngx.flush(true), >0 = background ngx.flush(false))
- Update EN/ZH docs: explicitly document 0 behavior vs >0 behavior
- Add TEST 7/8 to ai-proxy-flush.t: omit streaming_flush_interval_ms and
  verify the 10ms default activates the background flush thread

Co-authored-by: Copilot <[email protected]>

Replace route-setup + streaming request pair with a single check_schema
test that verifies the JSON Schema default (10) is injected into the conf
table when streaming_flush_interval_ms is omitted.

Co-authored-by: Copilot <[email protected]>