fix: recover interrupted model response streams

[minixalpha]

Summary

Recover interrupted model response streams instead of ending the run.

The SDK retries failures that happen before a response stream opens, but an
interruption during response-body iteration (ReadError, ReadTimeout, or
RemoteProtocolError) previously propagated out of the agent loop. SDK 1.5.0
uses httpx2, while the CLI only caught httpx.HTTPError; the two exception
families share no base class, so a mid-stream failure ended the session with no
retry.

Changes

• Bounded stream recovery (agent.py, transport.py): retry response-body
  read errors, read timeouts, and remote protocol errors at most twice, waiting
  1s and 2s. Retries reuse the last committed conversation, never execute tools
  from an interrupted attempt, and do not consume the turn budget. A 300-second
  window starting at the first interruption bounds new retry scheduling; an
  in-flight request keeps the SDK timeout and external run deadline.
• Transport compatibility: catch both httpx and httpx2 exception
  families through a small transport module, fixing the CLI's SDK 1.5.0 gap.
• Journal v3 (event_journal.py, event-journal-protocol-v3.md): record
  each attempt as a model.failed event with the original error, retry intent,
  duration, and the generation ID captured before reading the body. Readers
  still accept v1/v2; v1/v2 records containing model.failed are rejected.
• ATIF projection and cost reconciliation (atif.py, agent.py): project
  failed attempts as chronological incomplete steps, reconcile billing for
  interrupted generations, and mark token and cost totals partial when usage is
  missing.
• Docs: bilingual CLI reference section on interrupted responses, the v3
  protocol source plus English version, and the full 20-task live retest in the
  0.8.x dev notes (zh + en).
• CI: run the suite against Anthropic SDK 1.5.0 / httpx2 in addition to the
  locked dependencies.

Verification

Validation	Result
Complete suite, locked SDK / httpx	292 passed, 4 skipped
Complete suite, SDK 1.5.0 / httpx2 2.12.0	296 passed
Harbor adapter / ATIF compatibility tests	24 passed
Historical interrupted-response replay, SDK 1.5.0	4/4 recovered
tests/test_stream_recovery.py (run during the live experiment)	17 passed

Offline replay consumes the original captured response bytes through SDK 1.5.0,
raises the recorded transport error, and supplies a synthetic successful retry.
These replays make no external requests and add no task score.

Live experiment (20 tasks)

A full 20-task Terminal-Bench 2.1 run on this branch finished 20/20 trials:
11 reward 1, 6 reward 0, 3 unscored. All 576 model calls returned completely;
model.failed and interrupted HTTP response bodies were 0, so the recovery
branch was not exercised by real traffic. The run can only say "no stream
interruption occurred this run", not "live recovery succeeded"; the mechanism is
covered by the automated tests. Details are in docs/dev_notes/{zh-CN,en}/0.8.x.md
and the Git-ignored job record jobs/tb21-stream-recovery-pilot20-20260921-record/.

Notes

• No docs/research/ changes.
• No bilingual README changes, so no README sync is needed.
• Changelog: Fixed entry added under [Unreleased].