release: v2.1.0 — agent efficiency

[denfry]

Problem

On a 5.9k-file Java/Rust monorepo an agent using the skill spent as many tokens as an agent using Grep/Read alone, and more tool calls. Measuring why exposed real defects:

• refs/impact returned nothing for Class.method targets and mixed every same-named method for bare names; an empty answer came back as partial: false.
• Call edges bound across languages and onto other types' members; Kotlin had no call edges; Rust enum variants handled in match arms were invisible.
• Search packets were 7–10 KB of JSON whose snippets had no line numbers, so agents re-read files or grepped for lines before citing.
• Localisation catalogues and review diffs outranked code; a class lost its page slot to its own constructor.
• update parsed changed files serially (149 s vs 13 s for a full rebuild).

Solution

• Call edges store their receiver (index schema 4; the first update rebuilds once). Owner.member / module::fn targets for refs, impact, symbol, describe; untypeable same-named calls are listed as possible_call with coverage.partial: true.
• --compact agent output for search/explain/refs/impact: ranked locations with numbered matching lines, calling function and totals per refs site, build-module dependents in impact. refs --exclude-tests/--path, class cards in describe, exact-first symbol.
• Kotlin calls, Rust variant symbols and pattern references, parallel update.
• Three ablatable ranking signals (resource_priors, stem_match, question_pool_floor) and constructor folding.
• Skill rewritten around --compact and 40% shorter.

Verification

• 719 tests pass (coverage 86.5%), ruff, mypy, doctor --strict, check_versions, check_links, release_smoke.
• Retrieval eval over 342 queries on six query sets: MRR +0.010 (p=0.031), recall@10 +0.019 (p=0.017), nDCG@10 +0.011 (p=0.005), useful@budget +0.022 (p=0.008), no corpus regressing — tests/eval/results/2026-09-24-retrieval-2.1.0.txt.
• Agent pilot, three runs per arm: skill 83.5k tokens / 9 tool calls vs Grep/Read 103.0k / 12, all answers correct, ranges not overlapping — tests/eval/results/2026-09-24-agent-pilot.md (one private repo, five questions; scope stated there).

Compatibility

• Index schema 3 → 4: read commands keep working on an old index; the first update rebuilds it.
• JSON payloads gain fields (caller, target, receiver on refs sites, more_prefix_matches, modules, members/used_by/uses); nothing is removed.