[agent] Bound the OpenAI-compatible stream after finish_reason

[yusiwen]

Closes #176. Regression introduced by #175, found while delivering #174.

What changed

• agent/provider_openai.go now carries the same pair of bounds the Ollama provider does:
  openAIRequestTimeout (120 s) for one non-streaming request, and openAIIdleTimeout (2 min) for
  the silence between lines of a stream, reset by every line the scanner delivers. Both are package
  variables so a test can shrink them, and the cancel cause names the bound (errOpenAISilent), the
  same idiom as provider_ollama.go (issue agent: the Ollama provider has no request timeout, so a stalled endpoint hangs the run #1).
• A failed read returns an error — openai stream: … — instead of the partial answer with a nil
  error, so a cut stream cannot be mistaken for a finished one. That is also what the Ollama provider
  has done since issue agent: the Ollama provider has no request timeout, so a stalled endpoint hangs the run #1.
• openAIRequestError names the bound that fired and keeps errors.Is(err, context.Canceled) for a
  caller cancellation, which is how a TUI run renders as interrupted.
• chatStream takes the watchdog as an argument; the parser-level tests pass nil.
• Docs: CHANGELOG.md, CODEBASE.md.

Why

#175 had to read past finish_reason to reach the usage chunk, and removed the early return there
without replacing it. Measured against an endpoint that completes its answer and then holds the
connection open:

Endpoint ends with	before	after
data: [DONE]	2 ms, "Hi", err=nil	3 ms, "Hi", err=nil
nothing (held open)	1.502 s, "Hi", err=nil	1.5 s, resp=nil, err=openai stream: context deadline exceeded

The 1.5 s there is the caller's deadline in the repro; neither entry point passes one, so the real
wait was openAIRequestTimeout — the whole 120 s. The worse half is the second column: a truncated
stream was reported as a successful answer.

Rejected alternative: stopping at finish_reason once usage has been seen. An endpoint that never
reports usage would then return at the wrong point and cut the flush that follows the marker, which
is the defect #175 fixed.

Verification

• go test ./... -count=1 → 13 packages ok, 0 failed. go test ./... -count=1 -race → same.

• make staticcheck (v0.8.1) → exit 0. gofmt -l . → no output.

• The new gate was mutation-checked four ways:

  Mutation	Result
  watchdog never fires	the silence test times out at 25 s — that is the read blocking without the bound
  no per-line reset	a stream producing a chunk every 100ms failed: openai stream: stream went silent for 400ms
  a cut stream returns its partial text	a stream that went quiet after finish_reason returned success with "Hi"
  drop the caller-cancellation classification	on a cancelled context, error = http2: stream closed, want errors.Is(err, context.Canceled)

• The last row is why openAIRequestError has a direct unit test: on HTTP/1.1 the read error already
  wraps context.Canceled, so removing the helper's classification does not change the streaming
  subtest — the mutation survived until the helper itself was pinned with an HTTP/2-shaped error.

Honest scope

• No live provider call was made; the endpoints used are httptest servers.
• The idle bound is uniform, so a stream that goes quiet after a complete answer still waits the
  full idle bound (2 min) before failing. Distinguishing "the answer is complete but the sentinel
  never came" from "the model is still thinking" would need a shorter post-finish_reason bound,
  which this change deliberately does not add.
• A stream that closes without [DONE] is still accepted as complete, deliberately:
  OpenAI-compatible servers vary, and failing their working answers would cost more than the
  truncation it would catch. [DONE] still ends the read early.
• Behaviour change worth naming: a mid-stream caller cancellation now surfaces as an error rather
  than as partial content that looked like a finished answer. An interrupted run still reads as
  interrupted (pinned above), but the transcript difference is real.
• Reproduction harness: /tmp/dsh-repro-175 (scratch module, replace → this repository).