Extract shared invalidation from realtime layout commit dfeda25e6d. Retain original source ownership, detached-subtree cleanup, real splice ordering and structural-effect limits. Add selector/cache regressions and a normal-CLI before/after benchmark.
Large semantic dumps retain a JSON object graph with allocated field names
and string node references for every AX node before formatting the output.
Build compact typed nodes with numeric references and pass them directly
from the renderer to the CLI. Keep one AX traversal for both typed nodes
and CDP payloads, converting protocol nodes individually as they are built.
Preserve omitted/empty fields, field insertion order, list markers, and
frame-root attachment order. Render text iteratively and trim in place.
Add compatibility coverage and a standalone benchmark that stages both
executables on the same filesystem, alternates runs, records peak RSS and
wall time, and rejects changes to dump output bytes.
Against main at 5f5e1fff3 on a generated 40,000-row list (three interleaved
runs, medians):
- semantic_tree_text: 718.42 -> 390.96 MiB; 2.919 -> 2.204 s.
- semantic_tree: 815.71 -> 483.81 MiB; 2.484 -> 2.369 s.
- markdown control: 95.52 -> 95.59 MiB; 0.895 -> 0.913 s.
A generated 200,000-row list also completes with identical bytes (one run):
- semantic_tree_text: 3473.02 -> 1775.07 MiB; 13.245 -> 10.015 s.
- semantic_tree: 3803.45 -> 2137.26 MiB; 12.626 -> 10.397 s.
JSON/text comparisons also match the baseline on a mixed fixture containing
hidden/ignored nodes, form values, list markers, shadow DOM, and child frames.
These generated fixtures are independent of the external MOL-269 evidence.
Validation:
- cargo fmt --all
- cargo clippy --workspace --all-targets --all-features -- -D warnings
- cargo nextest run --no-fail-fast (19,885 passed, 17 skipped)
- cargo build --release -p moli
Refs: MOL-269
ldm0