[cristiangirlea]

Fixes #80.

A single document an operator runs the service from: deployment topologies (a sidecar, a shared worker behind TLS with per-client tokens, N instances behind a balancer with the rule that a job is answered only by the instance that holds it, the published images), sizing from the measured tables (one child against the 1536 MiB limit, the admission ceiling of capacity per retention, the connection budget, the drain deadline, the request limits), state (what a state directory keeps and what a restart reports), token layout and rotation, upgrades with a drain, monitoring (liveness versus readiness, the metrics to graph and alert on, the log schema), a failure-mode table, and the versioning and support policy. Every figure comes from docs/testing.md and every setting from the code.

Files: docs/operations.md (new), README.md (one link), AGENTS.md (the guide listed), CHANGELOG.md.

Checks on this head: the full deterministic sequence passed on the first commit (docs only; the second commit changes prose). Review: independent review by Codex (gpt-6.1-sol) against the code and the measured tables; its eight corrections (retention counted from completion, rotation without a no-refusal promise, model loading at child start rather than first job, the recorded versus computed throughput, unreleased wording, protocol compatibility limited to shared capabilities, graceful drain versus SIGKILL, best-effort logging) are in the second commit. Noted for the release PR: docs/testing.md line 93 still describes the earlier first-job model loading.