Code-switch span projection¶

openmed.clinical.languages.code_switch_projection extends source-aligned language routing with overlapping hypotheses. It never normalizes or changes source text. code_switch_windows(text, router=..., overlap=16) expands each routed run on both sides and clips at original character boundaries. The default router needs no model or network; routing confidence is not clinical accuracy.

Run a local detector separately on each text[window.start:window.end], then bind its local character spans to WindowEntity with the window ID, canonical label, score, and evidence digest. reconcile_code_switch_entities(..., text_length=...) projects spans to the original text. Exact span/label matches merge provenance, retain all language hypotheses, and record the largest declared detector score. Overlapping different labels or boundaries remain separate and all require review. Adjacent spans do not conflict. No score chooses a clinical meaning.

Reports contain offsets, controlled labels, language codes, scores and digests, never source snippets. Candidate status is only a detector hypothesis; this layer neither authorizes de-identification release nor replaces a clinical review gate. Invalid source/window/span bindings fail with controlled messages. Source size, window count, candidate count, overlap and conflict comparisons are bounded; dense ambiguous input fails rather than silently discarding candidates.