docs: crash-resume guide (docs/RESUME.md)

[prashar32]

Adds docs/RESUME.md — the full crash-resume model in one place, so the moat is documented as product rather than folklore.

It covers:

• The two halves — the daemon reloads runs with their already-spent budget automatically; the agent checkpoints its own progress and re-attaches with resume_run.
• What's restored — a table: spent tokens/dollars/loops, the budget, the cost ledger, and the last checkpoint payload; the time-budget clock restarts (it meters an active session, not downtime).
• Exact-once, even mid-step — the guarantee from fix(runs): exact-once resume across a mid-step crash #63: a partial step is rolled back so the loop and dollar budgets are charged exactly once.
• Idempotency, the one part that's on the caller — the interrupted step's work can re-run, so side-effecting tools must be safe to re-run. RiskKernel gates and audits them but can't make them idempotent.
• What can't be resumed — terminal (budget-halted / cancelled) runs; plus the runs resume / runs list CLI.
• Stability — the checkpoint format and resume surfaces are versioned (links COMPATIBILITY.md).

I also updated the kill-9-resume demo's crash-instant note to the now-exact-once behaviour and linked the guide from the README checkpoints row.

Docs only — no code change; verified every relative link and SDK-README anchor resolves.