chore(release): cut 0.1.34

[Jason Robert (jrob5756)]

Cuts release 0.1.34 (previous tag: v0.1.33).

Commit audit

All 9 commits in v0.1.33..main were audited — none omitted silently.

#	Commit	Subject	Disposition
1	4c98ab7	chore(deps): bump the github-actions group (#467)	Internal — CI workflows only
2	2dbb712	fix(mcp): support MCP 2.0 tool schemas (#420)	Fixed — already in changelog
3	3d9b071	feat(providers): native OpenAI provider on Pydantic AI (#421)	Added — already in changelog
4	2e5689d	test(fleet): wait on conditions instead of fixed sleeps (#478)	Internal — tests only
5	3c07139	feat(fleet): switch working directory for new runs (#479)	Added — entry added by this PR
6	d83e278	fix(fleet): Windows sharing-violation retries + dir picker (#487)	Fixed — already in changelog
7	134b735	fix(copilot): recover from a dead spawned runtime (#484)	Fixed — already in changelog
8	75eceda	fix(fleet): uncapped event-log streaming (#489)	Fixed — entry added by this PR
9	812a7dc	fix(providers): suppress idle watchdog during tool calls (#490)	Added + Fixed — already in changelog

Two user-visible changes had shipped without a changelog entry (#479 and #489);
both are added here. No existing entries were removed, and no empty category
headings remain.

Changelog summary for 0.1.34

Added

• Native openai provider on the shared Pydantic AI runtime, with MCP tools,
  structured output, interrupts, reasoning effort, and the full 0.0–2.0
  temperature range.
• Fleet TUI launch directory — d / ctrl+d picks the directory a new or
  resumed run is launched in and that relative workflow references resolve
  against.
• runtime.idle_timeout_seconds / runtime.max_idle_recovery_attempts
  (Copilot-only idle-watchdog tuning).

Changed

• pydantic-ai narrowed to pydantic-ai-slim[anthropic,openai].
• The reserved openai-agents provider name was removed from the schema,
  factory, registry and diagnostics.

Fixed

• The Copilot idle watchdog no longer fires during long-running tool calls,
  which previously clobbered an agent's structured output.
• MCP 2.0 tool discovery and structured tool results.
• Retry classification now covers pydantic-ai's ModelHTTPError /
  ModelAPIError, so 408/429/5xx retry on Claude and OpenAI.
• runtime.default_reasoning_effort is no longer dropped at run time.
• The Fleet Manager no longer loses the current step, token/cost totals, or
  topology on large or resumed runs' event logs.
• The Fleet TUI directory picker no longer clobbers its own prefill.
• Run-record removal no longer silently fails on Windows sharing violations.
• The Copilot provider recovers automatically from a dead nested runtime.

Validation

Run in an isolated chore/release-0.1.34 worktree off origin/main:

• make check — ruff check, ruff format --check, ty check all clean.
• uv run pytest -m "not real_api and not performance and not install_scripts"
  (the CI-equivalent selection) — 7652 passed, 46 skipped, 0 failed.
• make validate-examples — exit 0.
• uv lock — uv.lock changes only conductor-cli 0.1.33 → 0.1.34.

make test (which unlike CI does not exclude -m performance) reported one
failure, test_performance.py::TestForEachPerformance::test_100_item_array_with_max_concurrent_10.
That module is pytestmark = pytest.mark.performance and is excluded from CI's
required selection. It is a wall-clock deadline (~500ms baseline + 200ms
allowance) that lost its budget to load from the surrounding 7682-test run; the
test passes in isolation (2.26s) and the whole module passes standalone (22
passed in 12.54s). Not a product regression, and untouched by this diff, which
changes only version strings and the changelog.

Diff scope

CHANGELOG.md, pyproject.toml, uv.lock — nothing else.

After merge

Merging this PR will be followed by tagging the merge commit v0.1.34 and
pushing it, which triggers .github/workflows/release.yml to build and publish
the GitHub Release with its artifacts.