Harden HTML extraction and evidence file boundaries

[omar07ibrahim]

What changed

• replaces regex-based script/style stripping with a standard-library HTML parser
• keeps snapshot and evidence staging files owner-only throughout their private lifecycle
• adds regression coverage for spaced closing tags, executable snapshots, and publication modes
• adds a read-only deterministic refresh and adoption protocol for source-bound evidence

Verified evidence

• hosted refresh run 31382249540 regenerated and checked both real bundles
• all 13 non-manifest PNG, GIF, SVG, receipt, and CLI artifact identities remained byte-exact
• only the two provenance manifests changed, binding 72 handoff sources and 12 DeepThink sources
• final committed-bundle run 31382919187 verified both adopted bundles without checkout mutation

Verification

• complete locked CPython 3.12 offline suite
• Actions and Python CodeQL with zero PR alerts
• independent handoff and DeepThink evidence replay