fix: job --dry-run would_slice matches real-run slicing decision

[DLANSAMA]

Root cause

plate job/send <url> --dry-run --json predicted would_slice from the raw
URL-path extension via _predicted_url_download_extension, which returns None
when the URL path carries no file extension — the case for Printables model
pages (/model/12345-foo) and extension-less direct links
(/download?id=1). With no extension, the dry-run branch in
bambu_cli/job/orchestrate.py left would_slice at its False default.

The real run decides slicing from the downloaded file's extension
(orchestrate.py: if ext in SLICEABLE_EXTENSIONS). The downloader falls back
to .stl when it cannot infer an extension (_download_source_extension), and
the Printables resolver prefers STL/STEP — both sliceable. So the real run
sliced while the dry-run reported would_slice: false. The predictor and doer
computed "does this need slicing?" differently.

The fix

One shared slicing predicate, used by both paths:

• _ext_would_slice(ext) (bambu_cli/job/predict.py) — the single rule:
  ext in SLICEABLE_EXTENSIONS, applied to the single-suffix _file_extension
  so foo.gcode.3mf is print-ready .3mf on both sides. The real-run local,
  ZIP-member, and post-download branches in orchestrate.py all route through
  it.
• _predicted_download_slice_extension(url, args) — resolves the URL's
  predicted extension by reusing the downloader's own
  _download_source_extension, which lands the same .stl fallback the real
  download uses for extension-unknowable sources. --name is intentionally
  ignored because the real download overrides the --name extension with the
  URL/resolved-name extension (_download_filename_with_extension), so honoring
  it in the predictor would reintroduce a divergence.

would_slice stays a boolean; no JSON shape / docs/schemas change. For a
Printables page, remote_name remains null (unpredictable offline) — the fix
does not fabricate one.

Test coverage

• Extended the direct-URL dry-run matrix: .stl/.step/.obj → slice,
  .zip → extract, .3mf and .gcode.3mf → print-ready.
• test_dry_run_printables_model_url_predicts_slice — the regression: Printables
  model page now reports would_slice: true, remote_name: null.
• test_dry_run_extensionless_url_predicts_slice — ?id=1 predicts slicing.
• test_predictor_and_doer_share_slice_predicate — asserts the predictor's
  decision equals the doer's for the same resolved extension (drift guard).
• test_dry_run_local_gcode_3mf_is_print_ready_not_sliced — local .gcode.3mf.
• test_dry_run_local_zip_member_slice_matches_predicate — ZIP members
  .stl (slice) vs .3mf/.gcode.3mf (print-ready).
• Strengthened the Printables dry-run agent smoke to assert would_slice: true.

Sabotage-verified: reverting _predicted_download_slice_extension to the old
URL-path inference makes the three URL-prediction tests fail with the exact
would_slice False vs True divergence; restoring the fix makes them pass.

Gates (local)

• pytest -m "not live" with coverage: 1018 passed, 84.57% (floor 83)
• ruff check, ruff format --check: clean
• mypy -p bambu_cli: no issues (56 files)
• bandit -ll: exit 0
• privacy_smoke, python_compat_smoke, agent_cli_smoke: ok