Index measured label field overlap lookups with exact differential oracles

2026-09-13 05:32:55 -07:00
parent 4c9bc87e4e
commit 9b99e7c626
4 changed files with 531 additions and 19 deletions
@@ -0,0 +1,26 @@
+# Indexed field lookup experiment
+
+This unwired experiment extends source `4c9bc87e4e66bba573dd63b23df5ea039f85d0ca`. It does not adopt measured labels or field navigation in the renderer. This revision has **not been compiled or executed**. The baseline debug diagnostic reported roughly 2.4–6.6 seconds for a 4096-scalar plain field and 1.3–1.9 seconds for a 4096-scalar combining-heavy field; these are motivation, not results for this patch.
+
+`Label` now stores an overlap index built from the winning shaped buffer's clusters and the existing ordered, disjoint normalized scalar spans. For S spans, C clusters, and I actual span/cluster overlap memberships, index construction takes O(C log S + I) work and O(S + C + I) storage. A lookup visits the source spans and their indexed matches, then emits selected cluster rectangles in the original cluster/rectangle order. Hash membership replaces repeated linear rectangle searches. Signed zeros share a deduplication key, as with `Rectangle::PartialEq`, while the first original rectangle's bits remain unchanged. Hash iteration never determines output order.
+
+Per-cluster source end/start extrema preserve the old shared-cluster predicate, including whitespace and nonpainting spans. Source scalar overlap, truncation, missing-glyph and nonpainting precedence, partial/shared flags, invalid-range errors, session/font/zoom guards, and position/overflow rejection are unchanged. No new length limit is introduced. Index storage can be dense for intentionally overlapping synthetic clusters. Existing shaping and `field_clusters` extraction are unchanged, including that extractor's per-cluster linear rectangle deduplication; this is not a claim of globally linear shaping or preparation.
+
+The test-only `Label::original_field_layout` retains the baseline method body verbatim. Four new normal tests compare every `FieldLayout` field and every rectangle's exact component bits and position in the output:
+
+- Real Auto/Advanced shaping at two zooms and three widths, over every legal UTF-8 boundary pair, including scalar boundaries inside combining graphemes.
+- Overlapping, reordered, whitespace-spanning, missing, empty and reversed cluster ranges; partial retention and nonpainting/coverage precedence.
+- First-occurrence signed-zero bits, shared clusters and truncation.
+- Guard/error precedence, malformed fields, nonfinite positions and overflow.
+
+The existing ignored `field_lookup_4096_scalars_and_combining_heavy_full_coverage` diagnostic retains its two inputs and independent Paragraph coverage/rectangle oracle. It now times the original lookup separately, three indexed queries on the retained label, and three fresh-cache resolve-plus-lookup operations. Exact old/new results and independent complete rectangle sets are checked outside timed intervals. The original lookup precedes indexed samples; font caches are warm. These are debug experiment measurements, not a latency guarantee or a cold-process benchmark. Fresh resolution includes index construction so its cost cannot disappear into an unmeasured preprocessing step.
+
+The isolated harness should list 29 tests: 25 normal and four ignored. The production desktop does not register this module. Reuse the explicit dependency-bound harness recipe for the baseline, replacing its source binding with this revision and adding `scenario_label/lookup_tests.rs` to the input closure. Required execution is the normal suite plus these three ignored tests, each in a fresh process:
+
+```text
+scenario_label::field_tests::advanced_ligature_partial_field_uses_whole_shared_cluster_not_fraction
+scenario_label::field_tests::auto_ascii_to_ellipsis_advanced_transition_uses_final_buffer
+scenario_label::tests::actual_font_load_rejects_retained_label_and_old_session
+```
+
+Run the timing diagnostic separately with `--exact scenario_label::field_tests::field_lookup_4096_scalars_and_combining_heavy_full_coverage --ignored --nocapture --test-threads=1`. It remains optional functional-gate evidence; its exact full-coverage assertions are not waived. The new differential filter is `scenario_label::lookup_tests::` (four tests). The pinned-font checks retain their existing isolated-process contract. Other real-shaping differential cases use the current system default font and compare both methods on the same resolved buffer; they do not establish cross-machine glyph identity.