feat(deploy): make host cleanup a property of the host, not of deploying

docs/features/releases.mdmodified+44-0

Changes to docs/features/releases.md

		⋯95 linesShow the 95 lines above this hunk@@ -96,6 +96,50 @@ cloud release:pin <release-id> --remove
96	96
97	97	Retention candidates never include active, activating, awaiting-approval, pinned, or currently referenced rollback releases. The store keeps a configurable minimum per resource and exposes candidates rather than deleting provider artifacts implicitly.
98	98
	99	+## Host disk retention
	100	+
	101	+Release history is one thing a box accumulates; deploy machinery is another. Compute hosts also carry a content-addressed upload cache, staging tarballs, package caches, container images, and journals. These are reclaimed by a cleanup pass that runs **on a systemd timer as well as after each deploy** — a box that stops deploying keeps cleaning itself, and a deploy that fails no longer skips the pass that would have collected its leftovers.
	102	+
	103	+The pass reads root-filesystem usage before it does anything, and behaves differently by band:
	104	+
	105	+| Root usage | Behavior |
	106	+| --- | --- |
	107	+| Below `relaxedBelowPercent` (default 50%) | Reclaims deploy leftovers only. Caches that cost network to rebuild — the Bun download cache, container images, apt — are left alone. |
	108	+| Between the two thresholds | Full pass at the configured windows. |
	109	+| At or above `escalateAtPercent` (default 85%) | Full pass with every configurable window divided by `escalationFactor` (default 4). |
	110	+
	111	+Bounds that exist for correctness rather than retention never move with pressure: an in-flight upload belonging to a concurrent deploy on a shared box survives its hour however full the disk is.
	112	+
	113	+Defaults suit a typical box. A host packing many sites onto one disk is the case worth tuning:
	114	+
	115	+```ts
	116	+infrastructure: {
	117	+ compute: {
	118	+ cleanup: {
	119	+ // The upload cache mostly pays for itself within a single deploy — one
	120	+ // upload, N sites copying it locally seconds later. Keep it short.
	121	+ artifactMaxAgeMinutes: 720,
	122	+ escalateAtPercent: 80,
	123	+ onCalendar: 'daily',
	124	+ },
	125	+ },
	126	+}
	127	+```
	128	+
	129	+Components that write timestamped output to paths ts-cloud does not own register their own rules rather than accumulating unattended:
	130	+
	131	+```ts
	132	+cleanup: {
	133	+ rules: [
	134	+ { path: '/root/rpx-backups', pattern: 'cert-backup-*', maxAgeMinutes: 4320, directories: true },
	135	+ ],
	136	+}
	137	+```
	138	+
	139	+Each run prints root usage before and after, so the deploy log carries the trend. Set `timer: false` to keep the deploy-time pass and schedule the host's cleanup yourself.
	140	+
	141	+Rollback depth is separate and unchanged: `keepReleases` (default 4) governs how many past releases stay on the box, and cleanup never touches active or rollback releases.
	142	+
99	143	## Automation API
100	144
101	145	The versioned OpenAPI document and TypeScript client expose: