WP11: support realization-aware separator fitting

[IvanChernyshov]

Implements the approved WP11 workflow in #111: adds pyvoro2.inverse.fit_self_consistent_weights_from_separators with the exact agreed signature and re-exports the identical existing result class. The thin facade forwards to the concrete active engine, translating only max_outer_iter into existing options and selecting all candidates initially with history disabled.

Final-state inspection is Provisional across both imports of the shared class. Advanced engine/configuration, history, paths, counters and path-derived labels remain Experimental. Fixed fitting, solver mathematics, report v2, native qualification and all 53 advanced exports retain their existing behavior.

Public tests cover all five outer outcomes with independent numerical/geometric oracles, 2D/3D and periodic routes, mixed and bounded policy, output availability, strict validation, exact source binding and certificate failure propagation. Current references and guides document the three separate outcomes: outer termination, final inner status and final-state availability. Factual planning updates record WP11 as a candidate pending independent acceptance.

Validation at final head 194500338b5a782420fb438a967c0fe9a7f721d4:

• CI run 37310641186, attempt 1: all 49 jobs succeeded, including every child of the ten required FULL ∪ RUNTIME families, release validation and CI gate.
• Python 3.10–3.14: admitted GNU/Linux full suites, compatibility/runtime jobs, macOS arm64/x86_64 and Windows adapters. Ubuntu, runtime, AVX/FMA and rebuilt-sdist full suites each report 5,770 passed, 13 skipped; manylinux full suites each report 5,740 passed, 43 skipped. These suites overlap and are not summed as unique coverage.
• Notebook execution, tooling (880 passed, 7 skipped in the documentation job), strict MkDocs, flake8 and generated-file checks passed.
• Source-only focused selection: 104 passed, 37 deselected. Local controlled native qualification was refused by compiler-loader evidence, and local notebook execution could not start a kernel; native/full/tooling/notebook claims come from this exact-head CI.
• 72 artifacts inspected: 42 production qualification records and one sanitizer safety-only artifact. Archive, canonical manifest, effective-build, route/corpus, native payload, Python consumer, installed-process and wheel-record bindings were checked. The unchanged route validators also passed on all 43 retained production/safety evidence sets.
• Release matrix: 20 wheels and one sdist; each aggregate distribution matches its cohort hash. Release distributions, artifact 11349659070. Clean no-SciPy aggregate profile, artifact 11350071702.

Base/current dev: dadbddb9fe3496d96a3606c6808c6084f8b7d687. Final root tree: 2e3b762b51d3bdb6b7331f033d54d2bd3635480d. Canonical manifest SHA-256: e9ba7ad7be4c8a3c0a5121fb9ccda5ed0175ed53baa0be98cc89f20f32424965; aggregate source: 7e2ba7ff2d38025d063f997a95ad4c8d1c96708d387e7ea19fc318c200e0e81c. Manifest refresh was explicit measure/update/check; no automatic refresh or qualification bypass was introduced. Earlier-head runs are historical.

Implementation-side source review has no outstanding findings. Leave open for independent final review. Acceptance and integration remain maintainer decisions; this PR does not merge, close WP11, update #47 acceptance, accept Checkpoint C or start the later Phase C gate.