refactor: split converter/elements.py into a package with scoped conversion scale (v0.5.2)

[ShotaroKataoka]

Summary

Resolves the converter/elements.py known debt (principles.md): the 1519-line / 18-definition monolith is now a 4-module package with an enforced dependency DAG, and the converter's px scale is no longer process-global mutable state.

Internal refactoring only — no changes to public import paths, signatures, or JSON output (guarded by characterization tests; no CHANGELOG entry per versioning policy).

1. Characterization first (safety net)

• 41 tests pin the exact output of all 7 public extractors + _dispatch_shape, image/media file naming, the image counter, and pipeline JSON for standard 16:9 and non-standard 4:3 decks
• Fixtures are generated programmatically with python-pptx — no binary fixtures added
• Expected values were captured from the pre-split implementation and did not change through the refactoring

2. Scoped conversion scale (prerequisite for the split)

set_emu_per_px() patched a fixed string list of sys.modules entries to overwrite import-time copies of EMU_PER_PX. Splitting elements.py would have silently broken non-standard slide widths (new submodules missing from the patch list), and the process-global mutation already leaked across conversions (demonstrated: after converting a 4:3 deck, later direct extractor calls ran on a 4762.5 basis).

Replaced with ContextVar + conversion_scale() context manager + get_emu_per_px():

• always restores on exit (normal / exception / nested)
• thread-isolated (two concurrent conversions with different widths tested)
• AST guards forbid reintroducing EMU_PER_PX imports or sys.modules access in converter internals
• EMU_PER_PX = 6350 stays on the facade for import compatibility

3. Package split (mechanical move, bodies unchanged)

converter/elements/
├─ __init__.py   # compatibility facade (7 public extractors + _dispatch_shape)
├─ shapes.py     # name maps + line / freeform / auto-shape
├─ textbox.py    # textbox (delegates preset geometry to shapes)
├─ media.py      # video / picture / SVG / image fills / referenced images
└─ dispatch.py   # raw passthrough / type routing / group recursion

Dependency DAG (textbox→shapes, dispatch→{shapes,textbox,media}, facade re-export only) is enforced by an AST guard; submodules never import the facade.

4. Warning closure

• lxml FutureWarning fixed at the root: if grp_sp_pr → is not None (identical semantics)
• pytest promotes FutureWarning to error so it cannot regress
• defusedxml's own cElementTree DeprecationWarning ignored with a message/category/module-scoped filter only — the defuse_stdlib() security patch is untouched

Testing

• make all: 616 passed, 4 skipped, zero warnings (no existing test assertions modified)
• make smoke: OK (24 tools, 4 templates, persona served over real stdio)
• pytest tests/test_converter_elements.py tests/test_converter_robustness.py -W error::FutureWarning: 104 passed
• ASH scan: no findings from this change (17 pre-existing npm-audit findings in node_modules; bandit/semgrep PASSED)

Not in scope

Extractor redesign (Strategy registry / ExtractionContext), extract_textbox_element internal decomposition, and any schema/output changes — per the v0.5.x roadmap, one theme per release.