perf(extraction): compute a call's URL context once, and only when read

[DeusData]

Tonight's bench of the final release tree (10 LSP corpora, three interleaved runs each, plus the kernel) found csharp 10 % slower than this morning's main (135.5 s against 123.4 s, all three runs consistent).

A per-pass trace put the whole difference in parallel_extract: 24.0 s → 37.6 s. Bisecting over the merge batches pinned it to #2513's first commit (77bc492). A macOS sample diff showed the cost in handle_calls → cbm_service_pattern_match → match_qn → strstr.

Cause

extract_url_or_topic_arg computes url_context (is the callee an HTTP client or a global fetch?) once per positional argument of every call, in every language, through a linear strstr scan over all service patterns. Only the URL-folding functions read the answer, and they return NULL outside Python and JS/TS. The one other reader is the template-string branch.

Fix

A per-call url_context_t computes the answer at most once, and only when an argument reads it. The folding paths ask only in folding languages; the template-string branch asks for itself.

Proof (macOS arm64)

• C# extraction: parallel_extract 37.4 s → 24.2 s, wall 127–132 s → 119.7 s. Before Preserve topic identities while folding local URL constants #2513 it was 24.7 s / 121.9 s.
• Byte-identical graphs: the full node and edge sets, properties included, hash the same before and after on django (Python), typescript (TS) and csharp, HTTP_CALLS included (457 / 283 / 39). They also equal the current main bench binary's hashes.
• Tests: extraction, extraction_imports, pipeline, parallel, infrascan and edge_types_probe: 1009 passed.
• Lint: cppcheck 2.20 and clang-format are clean.