decimal-scaled

Use to verify the decimal_scaled repo still matches its architecture as defined — a read-only audit against the CLAUDE.md ARCHITECTURE CONSTITUTION (rules 1–6) and docs/ARCHITECTURE.md. Run it periodically, after any structural/perf change, or before a release. Covers the defect taxonomy (const-work-width / build-max leaks / per-tier pollution / layering / dead code / policy hygiene), the read-only protocol, what is legitimately NOT a violation, and a copy-paste launch skeleton for the auditor agent.

Use when the coordinator runs a PARALLEL multi-agent performance-recovery campaign on decimal-scaled — recovering many benched (bench-branch-compare) regressions at once by partitioning the work across worktree agents, one per policy/function area, with CPU-core-pinned concurrent microbenching. Covers the disjoint partition, int-tier-first holistic ordering, the core-pin scheme to keep concurrent benches under load, the per-agent loadout, and the coordinator's review→gate→merge→push→bbc→refresh-table loop. Builds on subagent-dispatch (per-agent rules), algorithim-optimiser (kernels), architectural-review (compliance), hunters (research lenses).

Use when writing or maintaining a DecimalSubject adapter or any golden-harness integration — adding a subject for a new decimal library to golden-competitors, fixing an existing adapter's limits/rounding/overflow declarations, or wiring a library against the decimal-scaled-golden golden set. Covers the typed-vs-erased choice, the Capabilities/FnSupport contract, value-aware limits, round-trip rules, panic capture, the gotchas hit building the seven competitor adapters, and how to validate a new subject.

Use when launching, re-dispatching, or handing over the decimal_scaled background "hunter" agents — the adversarial golden-input hunters (theoretical / competitor / internal lenses), the perf-opportunity hunter, the perf-research hunter, and algorithm candidate-writers. Covers the dispatch protocol, the hard rules (no cargo/bench by agents, cleanroom licensing, no background daemons), the come-back-resend loop, the manifests, the bench seams, and copy-paste launch-prompt skeletons.

Take a COMPLETED, verified policy-map (the table produced by `policy-mapper`) and wire it into ONE function's `policy::<fn>::select` — faithfully reproducing the map, validity-walled, at true crossovers — then prove the wired policy reproduces the map and golden-gate it. The apply half of the measure/apply split; `policy-mapper` measures, `policy-applier` wires.

Empirically MAP ONE function's algorithm policy by N-way-benchmarking every candidate algorithm × LimbSize across the full width/scale surface (+ edge input values) — producing the verified win-region map. DECIMAL sweeps the 5-point scale grid per width and bisects where winners change; INT brute-forces every width. Measure-only: the map is the deliverable; `policy-applier` wires it. One agent per policy; the coordinator verifies the map is complete.

Cut a decimal-scaled release. Points to RELEASING.md (the system of record) for the full process — versioning policy, release branch, PR-into-main + approval, mandatory docs + benchmark refresh, and publish steps. Use when preparing or performing any release.

Use when dispatching background/worktree subagents on the decimal-scaled project, or coordinating several in parallel. Bakes in the base-reset guard, no-AI-trailer rule, golden gate, keep-green discipline, no-push/merge protocol, check-in cadence, and surface-don't-proceed rule — all learned the hard way during the 0.5.0 multi-agent work.

Invoke when working with the `decimal-scaled` crate (const-generic base-10 fixed-point decimals D18/D38/D57/D76/D115/D153/D230/D307/D462/D616/D924/D1232 with 0.5-ULP correctly-rounded integer-only transcendentals). Covers when to pick the crate over alternatives, picking the right width and SCALE, construction (`d38!` / FromStr / `from_bits` / `from_int`), strict-vs-fast routing, rounding modes via `*_with(mode)`, the `Decimal` trait for width-generic code, the `DecimalConstants` trait for pi/tau/e/etc, and common anti-patterns. Apply when the user writes Rust code involving currency, prices, measurements, scientific values that must round-trip exactly through human-readable decimals, or anywhere they need deterministic bit-identical arithmetic across platforms.