feat: add llama.cpp runtime sizing editor

[sjug]

Summary

• Turn the verified llama.cpp runtime card into a typed automatic/custom sizing editor that seeds the exact persisted requested policy, previews aggregate context, and validates fixed gateway limits before submission.
• Follow the transactional relaunch through the existing provisioning task and log stream while identifying task generations by started_at, including fast terminal failures and later unrelated lifecycle operations.
• Protect multi-tab edits with pristine re-seeding, dirty-form staleness, explicit Reset behavior, confirmation, double-submit prevention, and fail-closed handling of ambiguous network responses.
• Add accessible responsive controls, operator and upgrade documentation, controller tests, and DOM-level regressions.

Safety and behavior

The browser performs only deterministic boundary validation and arithmetic preview. The gateway estimator and post-load runtime verification remain authoritative, and the node leaves routing while the disruptive relaunch runs.

An ambiguous lost POST response keeps the form locked until a new task generation appears or the operator reloads to reconcile current node and task state. The lifecycle lease rejects concurrent relaunch attempts.

Validation

• Python 3.12: 1,484 tests passed
• Python 3.13: 1,484 tests passed
• Coverage: 92.57% against the 92% floor
• mypy strict: 143 files, no issues
• Ruff check and format, ShellCheck, Node syntax checks, and git diff --check: clean

Post-merge validation

Run the agreed Phase 4 staging matrix, including the two-tab stale-form scenario and deep-JSON equality between the submitted custom policy and llamacpp_runtime.requested after relaunch.