refactor: use format.py constants for metadata record names in testing.py

[ayam04]

What changed

src/hflow/testing.py spelled format-owned metadata strings by hand at five sites ("episode/v1" and "robot_software_version"). It now imports and uses the constants that own those names, METADATA_RECORD_EPISODE and EPISODE_KEY_ROBOT_SOFTWARE_VERSION from hflow/format.py, mirroring the pattern already used in transform.py (its metadata handling at transform.py:444-446).

Sites changed (per issue #62):

• _video_episode_metadata: dict key + reserved_metadata_keys entry
• _write_video_episode (video fixture): writer.add_metadata(name=...)
• _write_episode_with_channels (channel fixture): writer.add_metadata(name=...) + metadata dict key

Runtime behavior is unchanged — the constants hold the same strings.

Validation (run in WSL2, since HFlow imports fcntl and native Windows cannot import it)

uv sync --locked --all-extras
uv run ruff check        # All checks passed!
uv run ruff format --check  # 96 files already formatted
uv run ty check          # All checks passed!
uv run pytest -q         # 371 passed, 3 skipped; 3 failed + 6 errors all in tests/test_ffmpeg.py

The tests/test_ffmpeg.py failures are environment-only on this machine (they need real ffmpeg/ffprobe binaries on PATH): I ran the same file on a clean main worktree and got the identical 3 failed, 18 passed, 1 skipped, 6 errors, so they are pre-existing and unrelated to this change.

Also verified the issue's DoD grep is now empty:

rg 'name="episode/v1"|"robot_software_version"' src/hflow/testing.py   # no output

Fixes #62

[kstonekuan]

Thank you @ayam04. Complete and correctly scoped, and you counted better than my issue did: the body listed four sites and there are five, since synthesize_episode's own metadata dict carries robot_software_version too. You caught it, which is the whole point of the change.

Also right to leave task, operator, and success as literals. EPISODE_KEY_ROBOT_SOFTWARE_VERSION is the only episode key with a constant in format.py, so converting the others would have meant inventing constants the format does not own yet. Matching what exists beats being tidy.

What I validated locally:

• Full quality gate is clean (ruff check, ruff format --check, ty check) and all 380 tests pass on your branch against current main.
• The definition-of-done grep comes back empty, and I confirmed zero remaining hardcoded occurrences of either string in the module.
• Generated a synthetic episode through synthesize_episode afterwards to confirm the fixtures still write a real, readable file rather than merely type-checking.

On the tests/test_ffmpeg.py failures: environment, not your change, and they pass here. Checking them against a clean main worktree before reporting is the right habit, and it is the second time you have done it. It genuinely saves me time.

Merging now.

One note I owe you, and it applies from here rather than as a complaint about this PR: the good first issue pool is best kept for people making their first contribution, and with #79 already merged you are past that. Nothing wrong with this one, and I would rather you had taken it than have it sit. But for the next one, the higher-leverage work is where you have already shown range: you did the layering analysis that shaped action_rate back in #24, and that kind of reasoning is worth more here than a mechanical refactor. Two live examples if either appeals, neither a starter issue: #86, where registering a check with required keyword arguments is accepted and then fails once per episode, and #85, where RuntimeConfig(api_port=True) slips through a range check because bool subclasses int. Never a rule, always an invitation.