fix(client): harden chat and agent SSE reliability

[coconut-yc]

Description

Fixes the streaming and response-fidelity defects reported in #1738 against
the main base ae9038732ad2. This PR is intentionally limited to the SDK
readers and the existing CLI/MCP accumulators; it does not change the CLI output
contract.

• Default timeout no longer severs SSE. Client.doRequestStream excludes the ordinary 30-second default from KnowledgeQAStream, ContinueStream, and AgentQAStreamWithRequest. Stream lifetime is governed by context cancellation. The public WithTimeout contract is preserved: when callers explicitly configure it, the same upper bound still applies to streaming calls.
• Large SSE lines parse. All three readers raise the bufio.Scanner per-line cap from 64 KiB to 4 MiB, covering reference events of hundreds of KiB while retaining a bounded allocation.
• Terminal error frames terminate. A response_type=error, done=true frame is delivered to the callback and then returned as an SDK error, even when the server leaves the connection open without complete or EOF.
• Agent accumulation waits for the run sentinel. AgentAccumulator now terminates only on response_type=complete via the new AgentResponseTypeComplete constant. Intermediate thinking/reflection/answer done:true markers no longer discard later tool events or the final answer.
• Reference identity survives unmarshal. SearchResult.KnowledgeBaseID, ParentChunkID, and SubChunkID preserve knowledge_base_id, parent_chunk_id, and sub_chunk_id from reference events.

The follow-up agent-output PR, #1741, is stacked on this fix and separately changes the CLI/MCP output contract. It adds bounded defaults, --reference, --verbose, human-readable text, and an explicit raw NDJSON mode.

Scope Boundary

The fix commit contains only:

• client/client.go, client/agent.go, client/session.go, client/knowledgebase.go, client/streaming_test.go
• the completion-sentinel hunks in cli/internal/sse/agent_accumulator.go and its test
• the corresponding complete-event fixtures in cli/cmd/session/ask_test.go and cli/internal/mcp/tools_test.go
• the streaming fixes under cli/CHANGELOG.md [Unreleased]

It does not contain projected events, buffered envelopes, --reference,
--verbose, or other output-contract changes.

Type of Change

• 🐛 Bug fix
• 📚 Documentation update (cli/CHANGELOG.md)
• 🧪 Test
• ✨ New feature
• 💥 Breaking change

Related Issue

Fixes #1738

Testing

New regression coverage verifies:

• the agent reader accepts 256 KiB and a payload just below the 4 MiB cap, while an over-cap payload fails cleanly;
• KnowledgeQAStream and ContinueStream each accept a 256 KiB reference event;
• the ordinary default timeout does not cut SSE, while an explicit WithTimeout remains enforced;
• terminal agent and knowledge-stream error frames return promptly even if the HTTP connection stays open;
• knowledge_base_id, parent_chunk_id, and sub_chunk_id survive JSON decoding;
• AgentAccumulator ignores intermediate done:true and finalizes on response_type=complete.

Validated locally:

(cd client && go test ./... && go vet ./...)
(cd cli && go test ./... && go vet ./...)
git diff --check

Checklist

• Relevant Go files are formatted and git diff --check passes
• Client and CLI tests/vet pass locally
• Self-reviewed the code
• Added regression tests for the changed behavior
• Updated the unreleased changelog
• No breaking output change; explicit WithTimeout behavior remains compatible

Screenshots / Recordings

N/A — SDK / CLI internals only.