starburst

• **starburst_map()/starburst_cluster() cost estimates match the backend used** — they forward launch_type/instance_type/use_spot, so an EC2 job no longer reports a Fargate price.
• **Cost estimates now use live AWS pricing** (Pricing API for On-Demand, EC2 spot-price history for Spot), cached per session with a built-in static-rate fallback when offline. Reported costs are **estimates** (measured runtime × current rate), not AWS billing figures; docs reworded accordingly.

New features

• **starburst_setup() now provisions the default EC2 capacity provider** (for c7g.xlarge) so the default EC2 backend works out of the box — previously the first starburst_map()/plan(starburst) run failed because no capacity provider existed. It is created at DesiredCapacity = 0 (no billable instances launch during setup). Pass setup_ec2 = FALSE to skip (Fargate-only users). (#37)
• **Custom EC2 instance types are now provisioned automatically on first use.** Selecting a non-default instance_type (e.g. c8a.xlarge) previously failed with an opaque "AutoScalingGroup not found" because only the default was provisioned. Now the capacity provider/ASG is created on demand at first launch (a one-time ~1–2 min step; no instances launched by provisioning), or on failure you get the exact starburst_setup_ec2(instance_types = "…") command to run.

Behavior changes

• **Detached session$collect() now returns an entry for every terminal task**, including failures. A failed task comes back as a structured failure (list(error = TRUE, message = …, task_id = …)) alongside successful results, instead of being silently omitted — matching the documented contract.
• Clarified session$cleanup(): by default it stops workers and marks the session terminated but **preserves S3 objects**; pass force = TRUE to also delete them. (Behavior unchanged; the docs were wrong and are now correct.)

Breaking changes

• **Removed the platform argument from starburst_map() and starburst_cluster().** It was ignored (architecture is inferred from instance_type: Graviton *g.* → ARM64, else x86_64), and its "X86_64" default even contradicted the ARM64 c7g.xlarge default. Drop it from calls; the backend picks the architecture from instance_type. (platform remains on starburst_estimate(), where it is functional.)
• **Renamed starburst_config(max_cost_per_job=) to max_hourly_cost.** The limit was always enforced as an hourly *rate* (USD/hour), not a total-job cap; the new name matches the behavior. Update any starburst_config(max_cost_per_job = ...) calls to max_hourly_cost = ....

Bug fixes

• **starburst_map() and starburst_cluster() now accept launch_type, instance_type, and use_spot** and forward them to the backend. Previously these functions had no backend arguments, so a call like starburst_map(x, f, launch_type = "FARGATE") silently passed launch_type to the mapped function .f (typically an "unused argument" error) instead of switching the backend — contradicting the 0.3.7 migration note. Backend selection now works uniformly across plan(starburst), starburst_map(), starburst_cluster(), and starburst_session().
• Fixed broken/misleading examples: the README detached-session example used a nonexistent detached = TRUE argument; S3-read examples used base read.csv()/ readRDS(url()) which cannot read s3://; and the API rate-limit example aggregated to 10× its stated global limit across workers.
• **Public base images are now built multi-arch (amd64 + arm64).** They were published amd64-only, so use_public_base = TRUE failed the multi-arch environment build with "no match for platform in manifest" for arm64. The build-base-images workflow now builds both architectures (matching the env build), so the public base path works. (#39)

Documentation

• Consistency sweep addressing an external documentation review: EC2 is now consistently presented as the default/recommended backend with Fargate as the optional alternative; the README and example console output match the current engine (ASCII status markers, one task per input element — no chunking step); the Getting Started tutorial leads with starburst_map(); a grouped reference index and an architecture diagram were added; and starburst_config() keys are now catalogued.
• Trimmed the Getting Started tutorial: the inline Monte Carlo / bootstrap / genomics snippets (which showed unbatched future_map(1:10000, …) and hand-written speedup claims) now link to the canonical, measured example articles and the Workload Shapes / Performance guides. The doc-consistency guard (tools/check-docs.R) now also flags naive future_map()/future_lapply() big-count calls, not just starburst_map().
• Removed the historical ROADMAP.md (a pre-1.0, Fargate-first implementation plan that also carried a stale "MIT" license note); the README Roadmap section is the single source of truth.