Finish the canvas/headless parity punch list

[d-chambers]

Description

Closes out the canvas/headless parity theme from the architecture review (following #47 and #48) — the three remaining small divergences:

Select — SelectionState.apply_select_params now accepts patch=None and builds the display-only state itself, replacing the widget's hand-built copy of that branch (duplicate basis derivation and field-by-field state assembly). Note: the review's original "payload writer lives only in the widget" finding had already been resolved in the meantime — patch_settings_payload lives on the model — so this closes the last remaining duplication.

Filter — the boundary-mode whitelists moved to nodes/filter.py as FilterMode/SavgolFilterMode Literals, and the params models now validate them: a headless SavgolFilterParams(mode="reflect") (which SciPy rejects) fails at validation with a clear message instead of surviving until the SciPy call. The widget's combo options are derived from the same Literals via get_args, so UI and validation cannot drift. Deliberately not moved: the legacy Gaussian-settings migration — it handles the pre-row-list flat-settings shape, which cannot reach the typed params layer at all (pre-_state workflows load as defaults per the documented storage break), so there is no headless path to diverge.

PlayAudio — the node was hollow: it declared time_scale/volume_percent params that nothing consumed while all the Qt-free DSP sat in the widget. Everything between a patch and device-ready PCM (shape validation, rate inference, robust-gain PCM normalization, default time scale, volume→dB mapping, output-rate clamping, block-wise resampling) now lives in nodes/playaudio.py; only the QAudioSink/QAudioFormat half remains in the widget. The params are now actually consumed: the new render_audition(patch, params) returns (prepared, output_rate_hz, pcm) headlessly.

CHANGELOG updated under Unreleased → Fixed.

Counterpart review: the Codex CLI is still rate-limited (verified again before this PR; resets Aug 8), so no counterpart review was run; the adversarial self-review record is in the untracked .scratch/review_parity_punchlist.md (covers the apply_select_params behavior-preservation check, why legacy blobs don't gain a new failure mode from the Literals, the deliberate non-move of the Gaussian migration, and a repo-wide stale-reference sweep after the PlayAudio move).

Checklist

I have (if applicable):

• referenced the GitHub issue this PR closes.
• documented new behavior with docstrings and a CHANGELOG entry.
• included tests (new tests/test_nodes/test_playaudio.py; mode-validation tests in test_filter_params.py; the PlayAudio DSP unit tests now target the node functions).
• run prek run --all-files and pytest tests (2114 passed, 44 skipped; all 10 hooks pass).
• for hot-path changes, run python scripts/bench_compare.py --baseline main. (not run — no hot path is touched: the moves are logic-identical relocations, the Literal validation replaces same-cost str validation, and PlayAudio DSP is not benchmark-covered)