refactor: isolate lossless cut planning

[anaschatz]

Problem

shorts_generator/local/clipper.py mixed pure lossless-cut command/key/path
planning with environment state, filesystem cache orchestration and FFmpeg
execution. The planning contract was difficult to review independently without
risking output behavior or existing test patch points.

Scope

• add 12 characterization tests before extraction
• extract only pure command, cache-key and cache-path planning into
  shorts_generator/local/lossless_cut_planning.py
• keep source identity I/O, cache hit/miss/touch/prune, temporary cleanup,
  logging and subprocess execution in clipper.py
• retain every existing clipper.py facade name, signature, default, return
  shape, log and exception behavior
• add the bounded engineering slice report

Base: 92988d82ae93845be42cc98cf0f0fd3600ef65a6

Head: 490304c9dd4d304df598f09293c8e61b41d04f23

Non-goals

No caption, highlight, Node, migration, API, infrastructure, provider, product
default, retry/fallback or video-output changes. No caller migration and no new
runtime dependency.

Dependency boundary

Before:

callers -> clipper.py
             |- command/key/path planning
             |- filesystem/cache orchestration
             `- subprocess execution

After:

callers -> clipper.py compatibility facade -> lossless_cut_planning.py
             |                              (standard-library pure leaf)
             |- filesystem/cache orchestration
             `- subprocess execution

The leaf does not import clipper.py, OpenCV, provider code or environment
configuration and performs no I/O, logging or process execution.

Measured production change

• production boundary: 130,371 -> 130,418 LOC, net +47 (about 0.036%)
• clipper.py: 7,609 -> 7,599 LOC, net -10
• new pure module: 57 LOC
• production diff: +70 / -23
• two production modules changed; no dependency added

The small positive delta is the explicit module interface plus compatibility
facade overhead. The planning implementation is not duplicated.

Exact regression evidence

• focused lossless/BF/highlights tests: 88 passed before the final purity check;
  12/12 dedicated characterization tests pass
• npm run lint: pass
• npm run build: pass
• npm test: 1,690 total; 1,683 pass; 7 explicit skips; 0 fail
• Python discovery: 513/513 pass (baseline 501; +12 characterization tests)
• isolated Python modules: 52/52 pass (baseline 51/51)
• HookGate quality: 96.2981 before and after; hard guardrails pass
• fixture fingerprint: 37ef626b9901ad2eeb1056d47954d88cdf2ede6430cc80e449bd5f9c7a91f9c9, unchanged
• metric fingerprint: d62ce934e038d9026263ce2be7efa5f0c6e3ffac8a583d7b23b609540a5f4fe9, unchanged
• git diff --check: pass
• no generated artifacts or import cycle added

Characterization freezes exact FFmpeg argv, timestamp/fps rounding, FFV1/PCM
mapping, stable cache key/path, source identity, spaces in paths, hit/miss,
disabled/zero-byte cache behavior, one execution attempt, logs, return values,
failure propagation and partial cleanup.

No deterministic video was rendered for a pixel/byte comparison, so this PR
does not claim pixel-identical output. It claims unchanged exact planning
contracts and full regression parity.

Next slice

Recommended separately: characterize and isolate import-time cache/environment
configuration. Do not start caption planning, Real-ESRGAN process planning or
highlights.py candidate discovery inside this PR.

[anaschatz]

Superseded by #3. The original PR retained a synthetic secret-like test string in its commit range, so the clean replacement was created from main without force-pushing or adding a gitleaks allowlist.