Code-index release measurements

Measured on 2026-10-11 for issue #9703. The measurement record contains all 58 trials, phase totals and invocation counts, graph and corpus digests, skipped paths, host load, memory readings, compiler settings, and binary/source identities. Repeated skipped-path lists are stored once under skippedFileInventories; each trial identifies its exact list through skipped_file_inventory.

Workload and boundary#

The frozen downstream corpus contains 8,463 indexed files plus one controlled Rust fixture, 88,173,033 source bytes, 752,584 graph nodes, 5,036,843 edges, and 441,301 resolved calls. Forty XML trace files are skipped; the record names them and distinguishes attempted reads from successful ingestion.

The benchmark host is an AMD Ryzen 5 3600X with 12 logical CPUs and about 60 GiB reported RAM. Both binaries use Rust 1.95.0 and the repository release profile: optimization level 3, thin LTO, and 16 codegen units. Their common source base is 059ef4c45271aed0f975c695082875724dfa55ae; runtime phase instrumentation is published separately in commit 7e21a3232ec04c375066282c8662780f108342d8. Hostnames and disposable absolute roots use public placeholders in the record; timings and source, corpus, and binary digests retain their measured values.

crates/harn-hostlib/examples/code_index_latency.rs calls the owning IndexState::build_from_root or CodeIndexCapability::restore_from_disk path and the registered query/reindex handlers. Cold means a new in-memory index; the operating-system file cache is not flushed. Restore readiness includes reconciliation and required persistence. Cold time includes final persistence. Resident edits exclude the initial restore and do not include an automatic snapshot write. Fixture preparation, compilation, lease waits, and graph fingerprinting are outside the reported operation time.

Body and declaration edits change the small fixture. The 64-file case forces actual parsing and graph replacement of selected call-heavy files by invalidating their stored content hashes, without changing their source bytes. The changed-HEAD case makes a real empty Git commit with unchanged source. These are controlled index workloads, not representative distributions of user edits. The benchmark requires an explicit disposable-corpus marker beside its /tmp root because preparation rewrites the fixture, snapshots, and Git HEAD.

The embedding host's optional Harn semantic reference resolver is excluded. These measurements do not establish whole-IDE startup latency or a comparison against another language server.

Release results#

Each row has three complete trials. Seconds are median [minimum, maximum].

Cold median time falls 62.9% (2.69 times faster); the 64-file restart falls 78.3% (4.60 times faster). The decisive phase is outgoing identifier references: cold median 97.886 to 1.898 seconds, and 64-file median 40.088 to 0.080 seconds. Cold parsing remains comparable at 25.666 versus 26.455 seconds. In the 64-file case, total reconciliation falls from 48.093 to 2.025 seconds. It also finishes before the existing 30-second checkpoint threshold, avoiding an intermediate whole-snapshot write; the checkpoint policy itself is unchanged.

Name lookup now has one owner that derives all-node, declaration, call-site, and per-file function candidates together. Reference expansion reads only name-addressable declarations and processes each distinct identifier once per file. Removal updates each affected name bucket once. Declaration eligibility, function visibility, and graph relationships retain their prior rules.

All 24 old/new comparisons match the complete normalized node/edge multiset, corpus facts, node/edge counts, and resolved-call count. Every restart/edit query returns the current fixture declaration. The normalized digest preserves duplicate facts while ignoring allocation-specific node IDs. Cold snapshots remain 583,840,101 bytes. Cold operation peak RSS is approximately 1.693 GiB before and 1.705 GiB after; the typed lookup costs about 13 MiB at this peak.

Restart control and limits#

The first matrix ran on a shared host with changing CI load. Five additional old/new unchanged-restart pairs used the preserved binaries, alternating which binary ran first, under one measurement lease. All ten graph/corpus comparisons match. Before median readiness is 7.902 seconds [7.845, 8.876]; after is 8.920 seconds [7.944, 12.147], an observed 12.9% median increase. The median paired graph-restoration delta is +0.025 seconds. Much larger differences occur in unchanged snapshot loading and trigram/word restoration, including a 1.9-second trigram-restoration difference in the final pair.

These trials do not establish warm-path equivalence or isolate the cause of the whole-restart increase. No warm restart speedup is claimed. They support the large reference-expansion improvement, whose phase reduction explains the cold gain and most of the 64-file gain. Resident fixture edits remain around 10 milliseconds; they do not establish large-file edit percentiles.

operation_peak_rss_kib is the benchmark process's Linux high-water mark read before fingerprinting. Resident cases include their untimed initial restore in that high-water mark. process_peak_rss_kib also includes the Cargo wrapper, compilation when present, and verification; it is not an operation-memory comparison. The index's estimated_bytes is an estimate, not observed RSS. Nested phase totals overlap and must not be added to their parents.

The next latency priority is snapshot/posting restoration, with a reserved-host baseline to separate runtime costs from shared-host variation before choosing a storage change. Parsing is the next cold-build target. Neither is rewritten in this change: reference expansion accounts for the largest measured avoidable cost, while storage and parsing need different ownership and verification decisions.

Behavior verification#

The owning regression parses 256 same-named calls and checks that the reference candidate scan contains only declarations, including after snapshot restoration, declaration replacement, and removal. Existing checks cover import visibility, ambiguous calls, reference order, and recovery from intermediate checkpoints. The rebuilt canonical CLI passes the code-librarian conformance cases for dependency views, member surfaces, and external callers with no skipped cases. Host-library tests, lint, and source/binary drift checks pass. No provider API spend is involved.