Current Work — Resume Bookmark

A transient bookmark for the next implementer: the immediate next-session entry point plus engine-internal items not fully captured elsewhere. The normative contracts live in docs/internal-spec/inference-engine.md and docs/adr/4-type-inference-engine.md; the forward-looking commitment envelope (release strategy + full backlog) lives in docs/ROADMAP.md; the released-version record is CHANGELOG.md. If this file disagrees with any of those, the spec / ADR / roadmap binds and this file is out of date.

Status

v0.1.15 released (2026-05-29). v0.1.16 prepared (2026-06-03) — version bumped + CHANGELOG.md § [0.1.16] sealed; bundle exec rake release is gated on explicit user authorisation and has not been run.

v0.1.16 lands the full plugin-contract interface-segregation + ergonomics suite (ADR-37/38/39/40), ADR-43 RBS-complete ancestor resolution + the make check-plugins gate, the v0.2.0 gate-1 executable evidence (external-plugin fixture + conformance / all-plugins-load / demos-run guards), RBS-robustness synthesis for malformed project signature_paths:, and the rigor-activerecord missing-schema memoization fix (Redmine −86% memory). Full detail is in CHANGELOG.md § [0.1.16]; do not re-summarise it here.

Headline realism numbers (measured at the v0.1.12 OSS-realism cut; still current — later cuts added onboarding / def.override-* / the plugin-contract suite rather than new full surveys):

Project	Scope	Before	After	Delta
Mastodon	app + lib	789	6	−99.2%
Redmine	full plugin set	163	79	−51%
GitLab FOSS	app/{controllers,mailers,workers,services}	~670	~140	~−79%

The 6 remaining Mastodon errors are unrelated to engine precision: 5 nil-receiver in test fixtures + 1 upstream ruby/rbs Resolv::DNS#getresources typeclass-narrowing gap (see docs/notes/20260528-rbs-upstream-pr-resolv-typeclass.md).

Next-session entry point

make verify is clean. The plugin-contract effort is released in v0.1.16; the two strategic levers left before the v0.2.0 cut are:

1. v0.2.0 gate 1 — the documented stability commitment for the external plugin contract. The executable evidence (external-plugin fixture + conformance / load / demo guards, all in CI) landed in v0.1.16; what remains is the documented “won’t break within 0.2.x” statement for the pinned plugin-contract namespaces (the drift spec already pins them). This needs explicit planning, not another incremental slice. See docs/ROADMAP.md § “v0.2.0 — first evaluation release”.
2. v0.1.17 — internal-structure review + performance tuning before the v0.2.0 cut: ADR-24 slice 4 (gated undefined-method on resolved closed-class self-calls), engine-internal precision uplifts surfaced by the v0.1.16 work, and perf follow-ups from the inference-budget survey / RIGOR_BUDGET_TRACE.

Everything else is demand-driven and lives in docs/ROADMAP.md § “Future cycles” (plugin-contract ergonomics follow-ons, dry-rb continuations, LSP capabilities, performance levers) — pull from there when a concrete need surfaces.

Reference reading

1. CHANGELOG.md § [0.1.16] — the plugin-contract suite + ADR-43; § [0.1.12] for the OSS-realism cycle.
2. docs/ROADMAP.md § “Release strategy — the road to v0.2.0” — what gates v0.2.0.
3. docs/internal-spec/public-api.md — the public-vs-internal stability boundary; cross-reference spec/rigor/public_api_drift_spec.rb before extending any pinned namespace.
4. docs/adr/2-extension-api.md — the plugin contract v0.2.0 must stabilise.

Queued tracks

Residual diagnostics

• Mastodon app + lib residue = 6 — 5 are genuine nil-chain bugs in spec/ fixtures (Mastodon-side, no Rigor work); 1 is the upstream ruby/rbs Resolv::DNS#getresources typeclass gap. The references/rbs branch widen-strscan-resolv-stdlib-sigs is staged — branch push + ruby/rbs PR creation are the user’s task.
• Redmine / GitLab FOSS residues are larger surfaces; each warrants its own survey cycle if the user wants to chase those numbers down.

Engine-internal (not survey-driven)

1. ADR-24 slice 4 — gated undefined-method / arity diagnostics on resolved closed-class self-calls. See “ADR-24 — implicit-self method-call resolution, remaining” below.
2. AR scope-body lambda self — scope :x, -> { select(...).group(...) } inside an instance lambda still needs the lambda’s self rebound to the model class. v0.1.12 closed implicit-self class-side resolution for ordinary method bodies; lambda bodies remain. Empirical case in docs/notes/20260523-mastodon-v4.5-regression-sweep-v0.1.9.md § “What is increasing” item 2 / ADR-26 territory.

Open engineering items

Engine-internal items the next implementer benefits from seeing directly. The full demand-driven backlog (editor mode, LSP capabilities, dry-rb continuations, ADR-10/13/16 follow-ups, performance levers, plugin-contract ergonomics follow-ons) lives in docs/ROADMAP.md § “Future cycles” and is the v0.2.x completion target. This section holds only items with engine-internal detail not captured there.

ADR-24 — implicit-self method-call resolution, remaining

• Slice 4 — gated undefined-method / arity diagnostics on resolved closed-class self-calls. Its own FP-evaluation gate (ADR-24 WD4) — a large new false-positive surface on metaprogramming-dense code, so v1 was deliberately precision-additive only.
• Non-Bot general adoption inside class bodies — resolved self-call return type is adopted ONLY when it is Bot. Unconditional adoption of precise non-Bot returns regressed rigor check lib by 16 diagnostics (pre-existing callee-return-inference imprecisions surfacing downstream); this follow-up needs callee-return inference precise enough that adopting precise types does not surface those imprecisions.

ADR-23 — rigor triage slice 4 plugin recognisers

Remaining: a Plugin hook letting plugins contribute their own recognisers (deferred). (receiver_type / method_name structured fields on Analysis::Diagnostic shipped in v0.1.8; the SKILL integration shipped with the v0.1.9 trio.)

Inference budgets — spec table is unwired; Layer 1 doc hygiene remains

The spec’s configurable budgets: table (docs/type-specification/inference-budgets.md) is normative-for-v1 but not wired — the only operative cutoffs are three hard-coded silent guards (recursion re-entry ≈ depth 1, ancestor walk 100, HKT fuel 64) plus ADR-10 budget_per_gem. Survey + the RIGOR_BUDGET_TRACE / RIGOR_HEAP_PROFILE / RIGOR_HEAP_TRACE probes landed 2026-06-03 (note docs/notes/20260603-inference-budget-reality-survey.md); the probes are reusable.

Layer 2 resolved, and it was not a budget. The large-app cost cliff was traced to 4.2 M retained Strings from one unmemoized failure in rigor-activerecord (schema_table_or_nil when db/schema.rb is missing) — fixed in v0.1.16 (Redmine 1518 MB / 173 s → 217 MB / 84 s). union_size was refuted as uncorrelated with memory. Budget wiring is now demand-deferred — no corpus project demonstrates a budget-shaped cost; if one ever does, re-run the 2a-style distribution probe first (ADR-41 WD3).

Layer 1 (demand-gated doc/spec hygiene, awaits ADR-41 acceptance): fix the docs/manual/03-configuration.md budget_per_gem description (it says “time budget in ms, default 1000”; really a method-def count, default 5000); reconcile recursion_depth (spec 5 vs the wired depth-1 termination guard — split “termination floor” from “precision-unroll depth”); add ancestor_walk (100) + hkt_fuel (64) rows to the documented table; author the missing user-facing budget explanation (placement TBD).

Stdlib RBS coverage-gap pattern

When an upstream ruby/rbs RBS gap is surfaced by a single internal Rigor call site, prefer (a’) an in-source # rigor:disable directive + load the library; when it surfaces across multiple call sites or in user-facing code, escalate to (b) a focused RBS overlay under Rigor’s own sig/, or (c) an upstream ruby/rbs fix. The references/rbs branch widen-strscan-resolv-stdlib-sigs (widens StringScanner#[], Resolv#initialize) is staged for an upstream PR — branch push + ruby/rbs PR creation are the user’s task.

Smaller queued items

• Sig-gen update_existing does not collapse sibling parent / child class blocks — merge_class resolves each candidate’s class_name independently, so flat-sibling layouts stay flat. Re-flowing an existing file into the nested layout is out of scope; workaround is to delete the target sig file and regenerate from scratch.
• Hash === expr case-equality narrowing (open3.rb:226 shape) — still open.
• In-memory Analysis::Runner.run_source(source:, path:, …) entry point — bypasses the per-call tmpdir + chdir in RunnerHelpers#analyze; ~5 % spec-suite win plus a clean public API for embedders (LSP / editor mode). Demand-driven.
• Sig-gen remaining gaps after --params=observed — attr_reader with ivars set from non-initialize sources (DB reads, config, side effects) still produce :untyped_return; fix is a hand-written sig annotation. Deep chains on untyped receivers → rbs collection install or ADR-10 source_inference:. Dynamic methods (define_method, DSL macros) → project plugin (escalation path A in the SKILL). Documented in skills/rigor-project-init/references/04-sig-uplift.md § “Step 5-d”.

Type-coverage uplift — line status (2026-05-23)

Phases 1–4 landed (String / Integer / Float / Comparable / Math / HashShape / Date / DateTime / Time). Remaining items, all release undetermined and tracked at full detail in docs/ROADMAP.md § “Future cycles” → “Type-language / engine”:

• Struct / Data value folding — ADR-worthy (needs two new carriers); Data.define is the likely better first target. Encoding value folding is a permanent exclusion.
• MathFolding result refinements — attaching range refinements (Math.exp → positive-float, Math.sqrt / hypot → non-negative-float) to the value-precise 28-function fold. Demand-driven.
• Hash rassoc shape handler — the one open low-priority Hash handler; value → [k, v] reverse lookup, foldable when every value is a Constant. Demand-driven.

Post-release follow-ups

• data/oss-sweep/mastodon-thresholds.json — refresh the stored thresholds against the ~6 baseline so the weekly OSS sweep gate is calibrated (the current file is uncalibrated, max_diagnostics: 999999).