Feature-class issues exceed the per-run budget; size work or add a per-issue budget signal

[Svaag]

The VPS launch-proof dogfood (AS215932/hyrule-cloud#28) showed the loop writes good, on-spec code but a feature-sized task exceeds the per-run budget (default 20 iterations / 45 min): every run ends needs_triage with partial work, and on the timer it re-attempts daily and self-heals away the partial work — no forward progress.

#10 added --max-iterations-per-run / --max-wall-clock-minutes-per-run overrides for one-off larger manual runs, but the hourly timer stays at the 45-min default, so any feature-class loop:approved issue queued to the timer will loop forever without finishing.

Options to decide

1. Sizing discipline: queue only budget-sized issues to the timer; decompose features into sub-issues (states+models / service+smoke / tests).
2. Per-issue budget signal: let an issue (e.g. a label like loop:budget-xl or a body field) raise the per-run budget for that change only.
3. Resume/continuation: allow a follow-up run to continue partial work for the same change instead of self-healing it away (interacts with the worktree self-heal from fix: loop substrate hardening — worktree self-heal + prod handoff allowlist #5).

Acceptance

• A feature-class loop:approved issue either completes to a draft PR or is explicitly deferred — it does not silently burn the daily budget with no progress.