Handle incomplete semantic analysis gracefully

The source-size limit has a structured diagnostic, but other semantic limits can invalidate or truncate internal state without a caller-visible reason. For example, exhausting MAX_FACTS produces an empty semantic match set that is indistinguishable from a clean file.

Work:

• Define a provider-neutral semantic diagnostic with a stable code, optional range, component/limit name, and observed and capped values. Keep syntax diagnostics distinct from incomplete semantic analysis in Rust reports, JSON output, and profiler totals.
• Thread one internal analysis-budget/status object through fact construction, resolution, constants, indexes, summaries, and object flow. Centralize defaults and expose smaller budgets only through test support.
• Retain sound facts already proven where possible, mark affected components incomplete, and prevent downstream absence from being treated as proof. Strict provenance chains crossing incomplete data must fail closed.
• Make fact exhaustion record one diagnostic without attempting a duplicate or out-of-range FactId emission.
• Add below/at/above tests for every limit, simultaneous exhaustions, ID conversion boundaries, deterministic diagnostics, and stable partial results.
• Classify the two diagnostics in the production sample. Preserve a minimal synthetic regression for any semantic exhaustion found there.

No files or tests are named. Start by tracing fact construction, resolution, constants, indexes, summaries, and object flow to identify the current limits and semantic status paths. Done means provider-neutral diagnostics, sound partial results, fail-closed provenance, centralized budgets, and the requested boundary, interaction, determinism, and regression tests.