Reduce member completion, hover and signature-help latency

[renkun-ken]

Completion, hover and signature help repeatedly replayed unrelated local assignments, re-inferred receivers and processed package defaults/delegation. Parse debounce also delayed completion after ordinary typing. This PR avoids that work while retaining static inference, structural-edit handling and production inference limits (250 ms / 20,000 nodes).

Changes

• Index accepted source scopes, function/default ranges, statement dependencies and binding histories. Requests use range lookup, follow needed definitions and resolve retained locals lazily; incomplete syntax keeps bounded recovery.
• Collect syntax names iteratively in C, including formal defaults and deduplicating symbols without per-node R allocation. Charge source/scope index contents to the parse-cache budget.
• Share successful receiver analysis across providers in a per-document cache bounded to 32 entries / 4 MiB. Invalidate on edits, accepted parses and semantic metadata changes; exclude exhausted/incomplete results.
• Skip member inference for ordinary symbols, disabled member completion and unrelated cached class metadata. Handle argument-less library() / require() and replenish scope candidates after shadow filtering.
• Capture only each deferred local's dependencies with request-local hashed lookup. Infer outer list/literal shapes for completion enrichment without populating full member-value memos; lexical shadowing and class-producing expressions retain full inference.
• Reuse package invocation summaries before defaults/delegation. Prepare return guarantees for short exported wrappers and methods using native-backed properties in the package worker, with deterministic staged probes bounded to 40,000 nodes and 1 MiB of summary accounting. Summaries retain class/type only; unknown variadic arguments, conditional fields, nested instance fields and replaced properties retain conservative inference.
• Remove debounce when promoting queued parses. Completion may reuse accepted scope for a narrowly checked final-line standalone identifier edit in an ordinary R script; structural edits, class sources, defaults and literate cells use current parsing.

Benchmarks

Installed R 4.6.1/macOS arm64 measurements with production request budgets. Pairwise comparisons use the same fixture and clear the relevant caches; results from different fixture/revision pairs should not be compared directly.

Source/scope indexes versus the first performance implementation (2303aaf), 1,000 locals, 15 sequential samples:

Workload	Before indexes	With indexes
Cold local list receiver	25 ms	1 ms
Cold receiver inside an R6 method	26 ms	2 ms
Cold R6 empty-prefix scope completion	85 ms	70 ms
Full parse, ordinary function / R6 declaration	74 / 74 ms	69 / 70 ms

Serialized parse size grows about 2–3%; the 1,000-binding index occupies about 0.5 MiB and is charged to the parse-cache budget.

Dependency capture and completion shape inference versus 9ce3a8b, alternating 25 requests per implementation, 1,000 locals:

Workload	Before follow-up	With follow-up
Cold R6 scope completion	73 ms	38 ms
Warm R6 scope completion	27 ms	13 ms
Ordinary scope completion, cold / warm	1 / 1 ms	1 / 1 ms

Full completion responses agree. Separate allocation profiling reduces cold R6 completion traffic from 68.6 MB to 17.1 MB; serialized parse size and full parse medians are unchanged.

Bounded package return preparation versus 7e3e7cd, separate installed R sessions, alternating 20 requests per implementation, Polars 1.16.0 long-chain fixture. Metadata is decoded before timing. Cold invocation clears both invocation and document receiver caches; warm invocation clears only the receiver cache.

Provider	Cold invocation before / after	Warm invocation before / after	Fully warm before / after
Completion	198.5 / 119.5 ms	126.5 / 29 ms	5 / 6 ms
Signature help	193 / 114 ms	125 / 26 ms	1 / 1 ms
Hover	186 / 109 ms	125.5 / 26 ms	1 / 1 ms

Cold p95 falls from 221.1 / 242.1 / 194.2 ms to 147.4 / 148.2 / 115.4 ms respectively. Responses agree after excluding the expected completion metadata-generation identifier change. Separate allocation profiling reduces cold completion traffic from 20.57 MB to 11.27 MB (three samples, median); only diagnostic profiling extends its deadline, and profiling is excluded from timed samples.

Polars worker preparation takes 4.824 / 5.768 seconds in this run. Summaries increase from 22 to 227, their object size from 64,096 to 660,352 bytes, and the serialized package snapshot from 4,913,841 to 5,067,466 bytes. Preparation remains outside the request path.

Validation

All 10 GitHub checks pass on aff6874: Linux/macOS/Windows R CMD checks, lint, stable/development S7 and all four coverage shards.

• Final installed completion, hover, signature, language-feature handler, member analysis/completion/providers/source, S4, class-scope and R6-symbol suites pass. Stable S7 has three expected capability skips; development-S7 member/class-scope suites pass.
• Regression coverage exercises indexed/fallback response equality, ordered effects, alias captures, nested/default scopes, accepted edits, serialization/cache accounting, lexical shadowing, completion limits, argument matching, variadic forwarding and conditional/nested/replaced fields. Preparation/inference never executes fixture inputs, defaults, constructors or methods.
• Member completion/providers and R6 coverage passes 791 assertions. Final member-completion/signature coverage passes 366 assertions, including the last field-preservation regressions. Native tests include defaults, ignored attributes, duplicates and deep syntax; an additional forced-GC probe passes. No new lints; git diff --check passes.
• Fixed concurrent coverage startup timeouts by isolating R6, Polars and signature fixtures in empty temporary workspaces. R6 awaits initial parsing before edit assertions; Polars has a coverage-specific metadata startup deadline. Immediate post-edit requests still disable retries. The original failing coverage shard subsequently passed; Cross-file signature tests explicitly await definition parsing through document-symbol replies. The final signature suite passes 70 assertions under coverage, and the complete local coverage shard 4 passes 657 assertions with one existing empty-test skip.

The PR includes production code and tests only. Benchmark scripts, CSV files and plans remain local.

[eitsupi]

Review by ChatGPT

I found two potential regressions that remain in the latest revision:

1. Argument-less library() / require() calls: member_package_scope_required() accesses item$expr[[2L]] without checking the call length. Both functions allow calls without arguments, so ordinary completion, hover, or signature help may fail. Please handle these calls and add regression tests.

2. Truncation before shadow filtering: scope_completion() now applies the candidate limit before keep_source(). When a name appears in both symbol and function candidates but one is subsequently filtered out (e.g., an enclosing function shadowed by a method-local variable), it can consume a slot and exclude another valid candidate. Please apply the limit after filtering, or replenish the candidates, and add a regression test.

Both issues should be addressed before merging.