fix(orchestrate): #1781 apr serve startup-ready timeout — configurable + size-aware

[noahgift]

Summary

Closes #1781. The hardcoded Duration::from_secs(30) in AprServeDriver::wait_for_ready blocked large MoE GGUFs (Qwen3-Coder-30B, 18.5 GB) from ever becoming ready — cold-cache load exceeds 30s; warm-cache is ~1s.

Root cause (5-whys)

1. Why did large MoE startups fail? → apr serve did not become ready within 30s
2. Why? → Cold-cache load of 18.5 GB Qwen3-MoE GGUF exceeds 30s
3. Why the 30s? → Hardcoded Duration::from_secs(30) at apr_serve.rs:143
4. Why hardcoded? → No env-var or model-size scaling
5. Why? → Designed for sub-2GB models that load in <5s

Fix

Two-axis: env override + size-aware default.

pub fn compute_ready_timeout_secs(
    model_size_bytes: Option<u64>,
    env_override: Option<&str>,
) -> u64 { ... }

Resolution order:

1. APR_SERVE_READY_TIMEOUT_S=N env var (operator escape hatch; clamped ≥1s)
2. Size-aware default: 30s baseline + 1s per 500 MB above 2 GB
3. Unknown size → 30s baseline

Per-model budgets under the size-aware default:

Model size	Budget
1 GB	30s (unchanged)
4 GB	34s
18 GB (Qwen3-Coder-30B)	62s
30 GB	87s

Error message updated:

apr serve did not become ready within Ns (override via APR_SERVE_READY_TIMEOUT_S)

Test plan

• 8 new unit tests in apr_serve_tests.rs (env precedence, clamping, invalid fallback, baseline, scaling, unknown size, env-when-size-unknown, real Qwen3-Coder-30B 18.5 GB)
• apr_serve module: 17 → 25 tests GREEN
• cargo check -p aprender-orchestrate clean
• cargo clippy -p aprender-orchestrate --tests -- -D warnings clean
• cargo fmt --check clean

Empirical evidence

• paiml/claude-code-parity-apr M260 dispatch produced 15/15 student-side driver_error with this timeout
• time apr serve run Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf against already-warm-cache model → ready in ~1s
• Companion-side pre-warm workaround at paiml/claude-code-parity-apr M262 (PR realizar: APR transformer loader lacks Q8/Q4 dequantization for attention weights #239) — this PR obsoletes the workaround for hosts that set APR_SERVE_READY_TIMEOUT_S

Doctrine

Toyota Way: fixed at root cause instead of accepting the 30s budget as a hard constraint.