refactor: isolate lossless cut configuration

[anaschatz]

Problem

shorts_generator/local/clipper.py owned the import-time parsing of the three lossless-cut cache environment variables alongside cache execution. This made the configuration contract harder to characterize and maintain inside the 7,599-line module.

Scope

• Add one standard-library-only pure leaf, lossless_cut_config.py.
• Preserve clipper.py as the compatibility facade and retain all three existing globals and monkeypatch targets.
• Freeze exact defaults, parsing, types, malformed-value exceptions, path handling, and import-time snapshot behavior with characterization tests.
• Add an engineering report for the slice.

Explicit non-goals

No changes to video output, FFmpeg arguments/codecs, timestamps, cache keys/schema/paths, hit/miss/pruning behavior, file touching, cleanup, logs, captions, highlights, HookGate, Real-ESRGAN, providers, Node services, APIs, migrations, infrastructure, authentication, or product features.

Dependency boundary

Before:

process environment
        |
        v
clipper.py parsing + facade globals + cache execution

After:

process environment mapping + LOCAL_CACHE_DIR
        |
        v
lossless_cut_config.py (pure parser, standard library only)
        |
        v
clipper.py facade globals + unchanged cache execution

Git identity and size

• Base: e0513092bc1ccd4ab40c821af84146084c6f5c6f
• Head: 735afb9aa615b599ed64c9f7cd858b6e812b331c
• Production delta: +44 net LOC
• clipper.py: 7,599 -> 7,590 LOC
• New production module: 53 LOC
• No new runtime dependency and no import cycle

Characterization evidence

Ten behavioral tests passed against the pre-extraction implementation. The extraction then added the pure-leaf and supplied-mapping boundary assertions, for 11 configuration tests total. They cover:

• exact defaults and exported types;
• enabled/disabled allowlist, case, whitespace, empty and unknown values;
• relative, absolute, empty, spaced, Unicode and home-expanded paths;
• binary GiB conversion, fractions, scientific notation, the 2 GiB floor and no upper cap;
• exact ValueError/OverflowError behavior;
• import-time snapshot and reload semantics;
• preserved direct clipper.py monkeypatch targets;
• absence of import-time directory creation.

Focused post-extraction checks:

• configuration plus lossless planning: 22/22
• BF editorial/cache patch-point suite: 48/48
• Python compile, import direction and git diff --check: pass

Full regression results

Gate	Baseline	Final
npm run lint	pass	pass
npm run build	pass	pass
npm test	1,683 pass / 7 skip / 0 fail	1,683 pass / 7 skip / 0 fail
Python discovery	513/513	524/524
Isolated Python modules	52/52	53/53
HookGate quality	96.2981, hard guards pass	unchanged
Fixture fingerprint	37ef626b9901ad2eeb1056d47954d88cdf2ede6430cc80e449bd5f9c7a91f9c9	unchanged
Metric fingerprint	d62ce934e038d9026263ce2be7efa5f0c6e3ffac8a583d7b23b609540a5f4fe9	unchanged

Limitations and warnings

• No fresh video fixture was rendered for this configuration-only refactor, so this PR does not claim fresh byte-identical or pixel-identical output.
• npm audit reports five pre-existing moderate vulnerabilities in transitive Hono adapter/Hono and PostCSS packages; there are no high or critical findings.
• The authoritative Node test runs required normal host execution because sandboxed Chrome launch/kill returned EPERM; both baseline and final host runs passed.

Next slice

Keep separate from this PR: characterization-first caption command/render planning extraction, retaining caption I/O, subprocess execution, logs, and facade patch points.