[jdsika]

Fixes #75. Also relevant to monarch-initiative/monarch-app#819 (slow CLI startup), since Monarch calls load_converter("merged.monarch").

Problem

Context.as_converter() relied on curies' default strict=True. That check (_get_duplicate_prefixes / _get_duplicate_uri_prefixes) compares every pair of records, so its cost grows quadratically. For each merged context (4,273 records) it takes 15–25 s, so every load_converter("merged") pays that on startup.

Change

• perf: Add _has_duplicates(), a linear scan for prefixes or URI prefixes (canonical or synonym) shared by two records.
  • When there are none, the converter is built with strict=False.
  • When there are some, strict=True is kept, so curies raises its usual DuplicatePrefixes / DuplicateURIPrefixes with the same message as before. The public API is unchanged.
• test: Run the two canonical-uniqueness integrity tests on merged, merged.monarch and merged.oak too (they pass). The alias tests still skip the merged contexts (Add tests for integrity of merged contexts #26). This is the follow-up offered in Contexts don't guarantee bi-directional mapping #11.
• docs: docs/intro.md built converters with Converter.from_extended_prefix_map(context.as_extended_prefix_map()), which is the slow strict path. It now uses ctxt.as_converter(). This also fixes examples that assigned ctxt but read context, and one line repeated three times.

Results

	before	after
load_converter("merged")	~14–18 s	~0.3 s
converters for all 11 bundled contexts	~85 s	~1 s
full test suite (local, w3id test excluded)	~27 s	~11 s, including the new test that builds all 11 converters

Verification

An independent adversarial review checked equivalence with curies' strict check:

• Code across versions: compared the relevant code in all 90 curies releases in the supported range (0.5.3–0.15.4). strict only ever controls these two exact-match checks, and from_extended_prefix_map passes strict= through in every version.
• Fuzzing: 20k random record sets plus hand-built edge cases, run on curies 0.15.4, 0.7.9 and 0.5.3 (Python 3.14 and 3.8). Zero disagreements with curies' check.
• Real data: for every bundled context, the converter built the new way is identical to the strict one: records, prefix_map, reverse_prefix_map, pattern_map, and 20k expand/compress probes on merged.
• Mutation testing: the new duplicate tests fail if the fallback, the URI-synonym check or the canonical-prefix check is removed.
• Lint: flake8 with the tox -e flake8 plugin set passes on Python 3.11.

Notes

• If a future curies release adds more checks under strict=True, they would be skipped for contexts that pass the linear scan. A linear duplicate check upstream in curies would be the long-term fix, and we're happy to propose one there.
• Three outputs in docs/intro.md were already stale on main and are left unchanged here:
  • FlyBase:FBgn123 no longer expands in the first example.
  • GEOGEO:1 now expands in the last two examples.
    These are data changes, not caused by this PR, since the converters are identical.

[jdsika]

The red check is not caused by this change.

• The only failure is tests/test_etl/test_w3id.py::TestW3id::test_w3id (KeyError: 'kgcl'). It also fails on current main since perma-id/w3id.org moved its namespaces under ids/ on 2026-09-04, and fix: read w3id namespaces from ids/ after upstream move #84 fixes it.
• The other 31 tests passed in the 3.12 job.
• The 3.8 and 3.11 jobs were cancelled by fail-fast. With test_w3id deselected, they pass locally: 3.8 and 3.11 both give 31 passed, 1 skipped, with the locked dependency versions (curies 0.7.9, bioregistry 0.11.10, pydantic 2.7.0).

I'll rebase this branch once #84 is merged so CI goes fully green.