refactor: isolate caption render planning

[anaschatz]

Problem

Per-frame caption cue selection, cue normalization, typography defaults, and Budget Friendly kinetic calculations were embedded in the 7.5k-line clipper.py, making these deterministic rules difficult to characterize independently.

Scope

• Adds one standard-library-only leaf: caption_render_planning.py.
• Extracts active cue selection, render-plan normalization, typography values, and kinetic-state calculation.
• Keeps _CaptionRenderer, drawing, _editorial_kinetic_state, caption_index, _render_elapsed, and all existing patch points in clipper.py.
• Adds characterization-first coverage and an engineering report.

Explicit non-goals: no cue grouping/timing changes; no font, layout, safe-box, coordinate, pixel, Pillow/OpenCV, FFmpeg, cache, provider, API, migration, infrastructure, or video-output changes.

Architecture

Before: frame loop -> embedded selection -> draw normalization -> kinetic math -> Pillow/OpenCV
After:  frame loop -> pure planner -> immutable results -> clipper facade -> Pillow/OpenCV

The new module imports only typing; it has no environment, filesystem, logging, subprocess, NumPy, OpenCV, Pillow, or clipper.py dependency.

Provenance and size

• Base: 33338d72cfa5de94370c0f2609fadf35f8057ae8
• Head: f300a78f30df6170b3d86505e5ffd6f1f26da72d
• clipper.py: 7,590 -> 7,468 LOC (-122)
• New planner: 218 LOC
• Production Python: 30,098 -> 30,194 LOC (+96)

The small positive total provides explicit immutable result contracts and a separately testable pure boundary while shrinking the monolithic facade.

Characterization-first evidence

• 9 behavioral tests passed against the original implementation before extraction.
• Final characterization suite: 15/15.
• Focused caption/profile/semantic/winner suite: 159/159.
• Existing deterministic caption ROI/pixel-preservation assertions pass.
• No full decoded-video comparison was added, so this PR does not claim independently proven pixel-identical complete videos.

Full regression results

• npm run lint: pass
• npm run build: pass
• npm test: 1,683 pass / 7 skip / 0 fail
• Python discovery: 539/539 pass (baseline 524/524)
• Isolated Python modules: 54/54 pass (baseline 53/53)
• HookGate quality: 96.2981, unchanged
• Hard guardrails: pass
• Fixture fingerprint: 37ef626b9901ad2eeb1056d47954d88cdf2ede6430cc80e449bd5f9c7a91f9c9, unchanged
• Metric fingerprint: d62ce934e038d9026263ce2be7efa5f0c6e3ffac8a583d7b23b609540a5f4fe9, unchanged

Limitations

The full browser suite needs normal host browser permissions; sandboxed Chrome exited with EPERM, while the same locked suite passed with browser permissions. npm audit reports five moderate transitive findings and no high/critical issue; dependency upgrades are outside this refactor.

Next slice

Characterize and isolate caption cue grouping/timing planning only, without combining font resolution, drawing, or render behavior changes.