[ata-sesli]

Summary

Implements #113. Small helpers that were copied between source files now have one definition each: the C ones in src/foundation/, the Zig ones in a new src/foundation/interop.zig. Where a copy behaved differently on purpose, that difference now lives in one named function built on the shared helper.

Stacked on #118. This PR's base is simplify/single-pipeline-path. Review and merge #116, #117 and #118 first; GitHub then retargets this PR to main.

Two commits share the C helpers, then the Zig helpers; two more clean up after #118 (see below). No behaviour change is intended. The output of indexing and queries should be identical.

Problem

Each copy was static, so a fix to one did not reach the others, and some had already drifted:

• The int-to-string log helpers had ring sizes of 1, 2, 4 or 16 depending on the file.
• heap_strdup in graph_buffer.c returned "{}" for NULL, while the copies in mcp.c and store.c returned NULL.

#118 deleted the sequential pipeline, and with it the read_file, build_import_map / free_import_map copies and most of the itoa_log copies the issue listed. This PR handles the rest.

C changes

Helper	Copies removed	Now
int-to-string for log arguments (itoa_log, itoa_buf, itoa_cx, itoa_k8s, pkgmap_itoa)	11	cbm_log_itoa in log.h. Thread-local ring of 16 buffers, the largest any copy had; the most any call site uses at once is 4.
heap_strdup	3	cbm_heap_strdup in compat.h (NULL → NULL). graph_buffer.c keeps its "{}" default as dup_or_empty_json, a one-line wrapper.
intptr_to_ptr	3	cbm_intptr_to_ptr, static inline in compat.h.
elapsed_ms	2	cbm_elapsed_ms in compat.h.
stat_mtime_ns	2	cbm_stat_mtime_ns in compat_fs.h.
path_basename, k8s_basename	3	The existing cbm_path_base (pass_configlink.c, pass_k8s.c, internal/cbm/helpers.c).
path_basename in registry.c	—	Kept as native_path_basename: cbm_path_base, plus splitting on \ on Windows.
file_ext	2	cbm_path_last_dot in str_util.h.
file_ext in store.c	—	Kept as file_ext_lower, because it lowercases into a buffer; it now finds the dot with cbm_path_last_dot.
ends_with	2	The existing cbm_str_ends_with. pass_githistory.c's copy took a precomputed length; every caller passed strlen(path).

Why a new cbm_path_last_dot instead of the existing cbm_path_ext: the file_ext copies in pass_similarity.c and pass_semantic_edges.c compare the text from the last . anywhere in the path. cbm_path_ext ignores dots in directory names and strips the dot, so for a path like conf.d/Makefile the result would change. That would change which functions count as having the same extension, and with it SIMILAR_TO and SEMANTICALLY_RELATED edges.

Also removed: the ring-size enums and the duplicate time constants (CBM_NS_PER_SEC, CBM_MS_PER_SEC, CBM_US_PER_SEC_F and others) that only these helpers used. pass_parallel.c used its old ring size PP_RING to size a padding field. That size now has its own constant, PP_RESOLVE_COUNTERS, with the same value of 5.

Zig changes

New src/foundation/interop.zig with the C-interop basics the ported modules each defined for themselves:

• logging: log, logInfo, logWarn (from cypher.zig, watcher.zig, userconfig.zig);
• formatting: bprint (from parser.zig and discover.zig; this is parser.zig's version, which also handles an empty buffer);
• allocation and C strings, formerly pub in parser.zig: mallocN, callocN, reallocN, safeReallocN, strFree, heapStrdup, sp, ciEq. heapStrdup is now the C cbm_heap_strdup, so the Zig copy is gone as well.

cypher.zig imports these from interop.zig instead of parser.zig. The cypher-specific helpers (ix, narrow, isScalarValueFunc) stay in parser.zig. interop.zig is added to CBM_ZIG_SRCS in Makefile.cbm; build.zig finds it through @import.

Follow-ups from #118

Two more commits clean up what #118 left behind.

test: drop sequential-path wording left after the pipeline unification. #118 removed the sequential pipeline and MIN_FILES_FOR_PARALLEL, but test names and comments still described a 50-file routing threshold and a sequential driver. The tests themselves still compare something useful, so they keep their assertions; only names, comments and diagnostic strings change.

• Renamed tests (old → new):
  • pipeline_issue54_*_sequential (8 tests) → pipeline_issue54_*_one_worker
  • pipeline_folded_topics_preserve_identity_sequential_parallel → ..._one_and_four_workers
  • pipeline_external_import_binding_sequential_parallel → ..._one_and_four_workers
  • pipeline_c_variants_on_sequential_and_parallel_paths_c1 → pipeline_c_variants_with_one_and_four_workers_c1
  • pipeline_rust_grouped_import_properties_valid_json_sequential → ..._one_worker
  • pipeline_typed_http_call_has_no_arg_url_twin_sequential and pipeline_capitalized_axios_import_is_http_client_sequential → ..._small
  • pipeline_seq_ts_cross_uses_shared_registry → pipeline_ts_cross_uses_shared_registry
  • handles_fiber_group_prefix_sequential_issue686 → handles_fiber_group_prefix_small_issue686
  • repro_seq_parallel_equivalence → repro_one_and_many_workers_equivalence
• Comments that said "pad past 50 files so the parallel path runs" now just say the fixture is padded. Comments that described the sequential resolver as current now describe it as deleted, or drop it.
• probe_hcl_infra_maps_edge was commented as an expected failure, because a one-file fixture used to skip the parallel path, the only one that emitted INFRA_MAPS. Every index now runs that path, and the assertion already passes. The comment now says so.
• A few src/ comments said the same: pass_lsp_cross.{c,h}, pipeline.h, pipeline_incremental.c, registry.c and graph_buffer.c.

Comments that describe the sequential pipeline as history ("before this fix, the sequential resolver…") are unchanged. So are uses of "sequential" that mean single-threaded, such as the gbuf merge.

style(watcher): satisfy zig fmt. One if in src/watcher/watcher.zig was not zig fmt-clean on main. The command format string is now a named const, so zig fmt keeps the brace on the if line.

Not changed

• itoa_imp in pass_importance.c formats a uint64_t, not an int, so it is not a copy of cbm_log_itoa.
• dupZ in watcher.zig returns a sentinel-terminated Zig pointer rather than a C pointer, so it was left as is.

Verification

• Full test suite with CBM_TEST_TIERS=all (macOS arm64): 6,449 passed, 0 failed, the same count as refactor(pipeline): run the parallel path for every index #118. The probes' UNEXPECTED PASS notes are identical to refactor(pipeline): run the parallel path for every index #118's.
• Full suite again after the refactor(pipeline): run the parallel path for every index #118 follow-ups: 6,450 passed, 1 failed. The failure was index_recovery_parallel_quarantines_crasher (an innocent file quarantined). That test depends on which files are in flight when a worker crashes, and this run shared the machine with another indexing process. The mcp suite rerun alone passed (176, including that test). The two earlier full runs passed it too, and these commits change only a comment in test_mcp.c. The total, 6,451, is the 6,449 above plus the two new unit tests. The probes' UNEXPECTED PASS notes are unchanged.
• New unit tests: log_itoa_keeps_sixteen_values (16 results from consecutive calls stay intact) and path_last_dot (including the directory-dot and NULL cases). Both pass; the log and str_util suites pass (65 tests). They were added after the full run above, so they are not in its count.
• Build sources: scripts/check-build-sources.sh passes.
• Format: clang-format-diff (clang-format 22.1.8) is clean on the changed lines. zig fmt --check (Zig 0.17.0) is clean on every changed Zig file, including watcher.zig.
• Other checks: git diff --check and tests/test_github_workflows.sh pass.

Diff (first two commits): 36 files, +563 / −713.

• Not checked: the Windows-only branches (native_path_basename and the Windows case of cbm_stat_mtime_ns) were compiled only by reading; there is no Windows build here. Each is a direct move of the old code.