test+fix: coverage push (rounds 1-2) and ray_vec_get_i64 DATE/TIME width fix

[ser-vasilich]

Summary

Two parallel-agent test-coverage rounds (10 agents total) bringing the suite from baseline upstream coverage to materially better, plus one real source bug surfaced and fixed along the way.

Numbers (post-merge):

• +12 test files (10 new RFL, 2 modified C-test in author's style)
• +1576 assertions
• Functions missed: 125 → 66 (~57 newly covered)
• Lines: 76.16% → 78.05%; Regions: 80.87% → 83.47%; Branches: 59.17% → 60.65%
• Per-file deltas: runtime.c 53→87%, fused_group.c 60→73%, group.c 56→65%, csv.c 71→77%

Tests:

• Debug build (-O0 -fsanitize=address,undefined): 2515 of 2517 PASS, 0 failed (3 runs, stable)
• Release build (-O3 -march=native -fassociative-math -ffp-contract=fast -fno-signed-zeros -fno-trapping-math): 2515 of 2517 PASS, 0 failed
• 2 skipped: 1 pre-existing (runtime/create_with_sym_oversized_file) + 1 new (public/vec_get_sym_id_w8 — auto-skips with clear message; W8 sym ID space below 256 is unreachable from public-API user-interns since builtin registration already consumes ~307 IDs)

Commits

e849f512 test: round1-2 RFL coverage push (10 new files)

• rfl/agg/rowform_topk.rfl — OP_GROUP_TOPK_ROWFORM / OP_GROUP_BOTK_ROWFORM (0% → covered)
• rfl/agg/rowform_maxmin.rfl — OP_GROUP_MAXMIN_ROWFORM (0% → covered)
• rfl/agg/rowform_sum_count.rfl — OP_GROUP_SUM_COUNT_ROWFORM (3..8 keys, 0% → covered)
• rfl/sort/fused_topn.rfl — top/bot over filtered vectors
• rfl/query/per_group_buf.rfl — nonagg_eval_per_group(_buf), const_str_expr_copy
• rfl/query/parallel_probe.rfl — idxbuf_hist_fn / idxbuf_scat_fn parallel row→gid probe
• rfl/temporal/extract.rfl — yyyy/mm/dd/hh/ss/minute/dow/doy (all 0% → covered, 8 helpers)
• rfl/agg/variance.rfl — var / var_pop / stddev / stddev_pop / dev
• rfl/io/csv_splayed.rfl — csv_splayed_writer_*, GUID writer, col_copy_str_pool roundtrip
• rfl/hof/wrappers.rfl — pmap / fold-left / scan-left dispatcher wrappers

All happy-path only (correct types / shapes); error/null/wrong-type branches deferred.

c421fac8 test: round1-2 C-level coverage push

Extends existing C test files (no spinoffs, matches Anton's style):

• test_fused_group.c (+575 lines): mk_combine_hist_fn / mk_combine_scat_fn / mk_combine_dedup_fn (multi-key fused-group 3-pass radix scatter combine, 0% → covered); fp_expr_const_str; fp_count_heap_* / fp_count_emit_keep_min (fused TOP-N count heap)
• test_sort.c (+425 lines): ray_top_fn / ray_bot_fn for I64/F64/SYM at K<N, K=1, K=N; topk_indices_cmp_single via SYM; msd_bucket_sort_fn + bucket_lsb_sort on N>1M
• test_public_api.c (+260 lines): ray_obj_type / ray_obj_attrs across all carrier types; ray_vec_get_i64/f64/sym_id across all element widths; ray_runtime_create_with_sym(_err); full eval interrupt API (ray_request_interrupt / ray_clear_interrupt / ray_interrupted and per-eval wrappers); ray_eval_get_nfo/set_nfo; ray_eval_set_restricted; ray_get_error_trace populated-after-error and cleared-on-next-eval contract

7444aff9 fix(runtime): ray_vec_get_i64 reads DATE/TIME with wrong width

Real source bug discovered while writing FFI introspection tests.

ray_vec_get_i64 dispatched RAY_DATE and RAY_TIME through the same 8-byte cast as RAY_I64 / RAY_TIMESTAMP. But ray_type_sizes in src/core/types.c:37-38 declares both DATE and TIME as 4-byte elements:

• idx = 0 → reads 8 bytes, upper 32 bits captured from the adjacent element → returns garbage
• idx ≥ 1 → strides past row boundary → out-of-bounds read

Fix: split DATE / TIME off the int64 path; read them as int32_t alongside RAY_I32. RAY_TIMESTAMP stays on the int64 path (it's genuinely 8 bytes).

TDD: two tests in test_public_api.c (test_public_vec_get_i64_date, test_public_vec_get_i64_time) covering known DATE / TIME values fail before the fix and pass after.

Test plan

• make clean && make test (debug, ASan+UBSan): 0 failed, 3 consecutive stable runs
• Release build (-O3 -march=native -fassociative-math …): 0 failed
• All upstream sentinel-migration-finish tests intact (initial phantom failures were stale .o files from pre-pull state — clean rebuild eliminated them)

Out of scope (follow-up issues, not in this PR)

• Round 3 worktrees (9 more test files, ~700 assertions) — kept for a separate PR
• Dead code observations surfaced by coverage runs: fused_topn_fn + helpers in sort.c, aggr_unary_per_group_buf + aggr_med_per_group_buf in query.c, normalize_columns in datalog.c
• Structurally unreachable safety branch in collection.c (hashset_grow — over-provisioned init keeps load factor < 0.5)

🤖 Generated with Claude Code