[SizzleUnrlsd]

Summary

Opens Phase 5 of docs/architecture.md with §16 Security Model Registry and §17 Global Taint Engine.

• Acceptance tests first (test: define security models and taint engine behavior), implementation follows.
• taint/models.py: TaintKind bitset (SQL, COMMAND, HTML, PATH, SSRF, CODE, joined with |), Source, Sink and Sanitizer models keyed by canonical symbol, a mutable SecurityModelRegistry that plugins fill and freeze into an immutable ModelTable, exposed to the manager as the taint.models input.
• AnalysisManager.provide: a generic way to supply the result of a module-level analysis the engine computes elsewhere, with MissingInputError when such an input is requested before being provided. This keeps the manager free of any security-specific knowledge.
• taint/engine.py: one shared TaintAnalysis (taint.flows) per function, a forward data-flow problem over the SSA form. Source symbols and their call results carry taint; arithmetic, attribute and item access, iteration, phis and unknown calls propagate the union; sanitizers clear their kinds; comparisons and literals carry nothing. Every tainted argument reaching a sink whose kinds it still carries becomes a TaintFlow with its source, sink, kinds, argument and call location, in block then instruction order.

The engine is the second client of the data-flow solver after constant propagation, and the first consumer of provide.

Not in this PR: findings and reporters, and the detectors that turn flows into findings (next two Phase 5 items). Method sinks reached through attributes on objects (cursor.execute) wait on framework models and the call graph.

Closes #19