[Sunrisepeak]

Cold course startup can spend more than 120 seconds preparing a toolchain before the Provider executable exists. Give initial describe an absolute 900-second deadline, show preparation before session discovery, and retain independent stdout/stderr diagnostics. Subsequent Provider calls keep their existing idle limit.

Propagate startup, enumeration and check transport failures to structured error events and a nonzero CLI exit. Kill contained child process trees on timeout/cancellation, reject partial successes followed by child failure, and install downloaded course workspace dependencies before reporting installation success.

Validation: local build, portable startup/failure injection and session tests; CI covers Linux static distribution, macOS ARM64 and Windows 2022/2025. CI also produces candidate binaries with source SHA and SHA256 for d2mcpp integration before release. Documentation describes the configuration and event contract.

Self-review completed:

• Initial deadline is monotonic and never refreshed by UI ticks or continuous diagnostics; bounded drain batches prevent output floods from starving timeout checks.
• Only stdout is protocol input; stderr JSON cannot spoof a describe event. Raw logs are separate and machine output remains valid NDJSON.
• Partial describe/exercises/verdict followed by child failure cannot advance progress. Windows containment fails closed; POSIX timeout/cancellation reaps the child and kills descendants.
• Configuration round-trip preserves startup/idle limits. Both print and TUI preparation/error rendering were exercised locally.

Validation evidence: all six native CI jobs passed in https://github.com/d2learn/d2x/actions/runs/39441449900 (Linux static, macOS 14/15 ARM64, Windows 2022/2025, real Linux course smoke). Portable startup tests include installation/dependency failure and descendant cleanup; existing protocol e2e and 28 session assertions pass. Local Ubuntu 20.04 ordinary-user container reached a cold session in 99.2s, passed save/advance/resume and all 52 reference answers using the explicit static candidate. Full cross-repository candidate matrix is being checked before stable release.

The complete d2mcpp candidate matrix is now green: https://github.com/mcpp-community/d2mcpp/actions/runs/39443060283 (10 first-use jobs, separate README/cold, Linux without a compiler, macOS ARM64, Windows 2022/2025 including no MSVC). Artifact checkout SHA 3f56f74 and binary SHA256 were asserted in every job. Candidate testing explicitly overrides the installed stable binary; post-release course CI will verify the actual xlings-installed release too.