Shared progressive inventory controller

Shared ownership

common/web/progressive-loader.js exports PrintMasterProgressive in browsers and CommonJS exports in tests. It has no framework or page-specific dependency. common/web/shared.go embeds ProgressiveLoaderJS; Server serves it at /static/progressive-loader.js, before its application script. The Agent can reuse the controller by adding its own asset route/script and transport adapters; Agent integration is not implemented here.

createLoader accepts injected fetchIndex(context), fetchRows(keys, context), fetchMetrics(keys, context), getKey(item) and onChange(event). Context contains signal and generation; adapters should pass the abort signal into transport. Change events contain stage type, affected keys and generation. Transport failures must reject; successful stages must return arrays, including empty arrays for no authorized matches.

Lifecycle and demand

1. load() resets/aborts prior work, then requests only the index. getIndexState() distinguishes idle, loading, ready and error. Empty successful index is ready.
2. Render identity previews from getIndex(); apply global metadata filters/sorts before choosing DOM pages.
3. observeViewport({elements, sentinel, getKey, onDemand, onMore}) owns viewport demand and infinite-page observation. Default overscan is 200 pixels. Pass current DOM rows/cards and optional sentinel. Disconnect the returned observer before changing views/pages. Keep a native Load more button as keyboard/no-observer fallback; without IntersectionObserver the helper demands the first 30 supplied elements.
4. setDemand(keys) deduplicates authorized viewport keys and replaces queued row demand. It does not eagerly hydrate the full index.
5. Render completed rows using getRow(key). Only after hydrated DOM paint, call markRendered(keys) to enable lazy metrics. Server uses animation-frame boundaries and DOM hydration markers. getMetrics(key) and getState(stage, key) drive pending, ready, missing and error cells.
6. Refresh calls load(); reset() can abort/clear without another index request. Generation checks ignore stale completion even if an injected adapter ignores its abort signal.

Both stages share a request scheduler: rows take priority over queued metrics. Default batch size is 30 (clamped to 1–100); default concurrency is two (clamped to 1–4). Existing pending/ready keys deduplicate; unsolicited response keys are discarded. Partial responses mark absent keys missing without a fetch loop. Queued offscreen demand is dropped; already-running batches may finish.

retry(keys) retries errors explicitly, default maximum two attempts per key/stage (configurable up to three). It does not retry missing results or successfully cached rows. Once exhausted, refresh is required. Index retry is a new load(), not an automatic loop.

Global supply checks

requestMetrics(keys) is an explicit exception to viewport metrics demand, used only when a global supply filter/sort requires snapshots. It scans in bounded batches without fetching full rows. clearMetricQueue() ends that scan and retains only queued visible metrics. It cannot abort an individual already-running batch; reset aborts the whole generation.

Retain a lightweight per-index-key supply classification separately from the bounded snapshot cache. An evicted snapshot must not erase a completed global classification and silently exclude a distant match. Distinguish pending/error from genuinely unknown/no snapshot. If results are incomplete, say so rather than displaying a definitive empty inventory.

Server metrics adapters consume the exact reviewed snapshot projection: serial, agent ID, timestamp, optional page count/color pages/mono pages/scan count/toner levels (id is zero). Snapshot omission is not a zero counter. Zero is a real value. Inventory raw_data can include stale toners; the progressive list strips supply and snapshot-counter fallbacks until the queried snapshot is available. Copy/print/fax/duplex counters are not supplied by this projection and remain absent.

Memory and scope

• Defaults: 300 full rows plus 300 snapshots, separate LRU-like insertion/touch caches; cacheLimit cannot be smaller than batch size.
• Full lightweight index, authorized-key set, stage state/attempts, rendered-key set and optional supply bands are O(authorized devices), retained until reset.
• Server preview DOM accumulates 30-row pages while scrolling; not full virtualization. Hydrated values are not copied into the global index array, so cache eviction actually releases the full objects.
• Revisited evicted ready entries can refetch. No TTL/background metrics polling is introduced; refresh starts a new snapshot generation.
• Index metadata supports global search/sort; agent names and tenant memberships enrich from independent directories. Do not pretend MAC/firmware or unqueried counters exist globally in the index.
• Browser caches are not authorization boundaries. Every index, row and metrics request is authenticated/authorized by the Server. Never reuse this Server adapter for Agent-local routes or machine credentials without reviewing those different contracts.

See reviewed backend contract and user behavior. The Server UI integration and reusable controller were reviewed at source revision 11f502c. This frontend slice changes no OpenAPI schema, HTTP route version, machine protocol or product version.

Validation

Jest covers delayed/partial/error stages, explicit retry caps, deduplication, reset/stale completion, cache eviction, batch/concurrency limits and shared observation. Playwright serves actual Server HTML/scripts with contract-shaped fixtures, delayed index/rows/metrics/directories, no legacy list requests, bounded first-screen demand, scroll expansion, distant metadata/supply matches, cache-preserving view changes, zero counts, loading/error/empty states and keyboard controls. Run all six configured projects (Chromium desktop/small/mobile, Firefox desktop, WebKit desktop/mobile), including tenancy, table-customizer and context-menu suites. Go static-asset tests verify embed/route/MIME/cache wiring.