What the glue is doing

Status: Accepted · Audience: Both · Area: N — Companion

---

Purpose

Area H is the automation people bolt onto a media stack by hand — cross-seeding, announce-driven grabbing, quality-profile sync, library cleanup, playback statistics — and K1 is the metrics pipeline beside it. Five of the six share one rule: they prove the wiring rather than trusting it, and a configuration that has not proven itself is unproven, never ready. Cleanup has a gate in its place — a dry run, then a confirmation.

That rule is the one a small screen loses first. An automation that is running looks, from a phone, exactly like one that is working, and every one of these fails by looking fine. So what this page asks of the app is mostly the same thing each time: show the proof, keep unproven apart from ready, and offer the acts each feature makes reachable without a terminal.

Queue self-healing (H5) is not here: N16 already shows what it would do and did (N16-R8 to N16-R11), and keeps its acts off the app (N16-R12).

Behaviour

Unproven is its own answer, everywhere on this page

Five of these end their wiring with a proof against real data — a match for a real local item, a real announce line through a test filter, a read-back of the scores that were written, a known recent play queried back, a series present in the store. Where the proof has not held, the feature is unproven, and the app shows that as its own state. Running is never shown in its place.

Cross-seeding: a match is the data, not the name

A candidate is a match only when the files, sizes and pieces agree (H1). A near-miss — an extra folder, a stray small file — is a near-miss, and a partial match is partial with its missing pieces named. Neither is added as a seed, and neither is counted as one here. An indexer that failed during a pass is named as failed for that pass while the others go on; a client credential that failed its proof halts additions, and the app says additions are halted rather than that nothing was found.

The operator can start a search, list what it matched and run the proof from here.

Announce-driven grabbing: a quiet channel and a dead one are different

Each announce channel has a connection state and an idle time (H2). A disconnected watcher looks like a quiet night until a missed release proves otherwise, so the app shows both, per channel, and never shows a disconnected channel as watching.

A filter that has matched nothing over its window is a dead rule until shown otherwise. A filter at its grab cap is holding, not failing. A match whose destination was unreachable is held, not dropped. Each is shown as itself. The operator can reload filters, replay an announce line through them, and run the proof from here.

Quality-profile sync: the preview is the decision, and the read-back is the result

A sync is previewed before it is applied — every create, update and, where pruning is on, every removal, with the score it changes (H3). The app shows the preview and applies only the plan that was previewed. Afterwards, success is what the read-back confirmed, instance by instance: a write the service accepted and the read-back did not find is a failure, and another tool rewriting the same instance is a conflict, not a reason to apply again.

Library cleanup: the evidence travels with the candidate

A cleanup candidate is chosen from what the household watched and asked for, and carries the rule that chose it, the room it gives back and the evidence behind it (H6). The app shows all three with each candidate, and shows only the evidence the stack gave: a child’s viewing is not evidence the stack gives, and the app has no other source. A series is shown as its members, each with its own evidence, never as one row to delete. A confirmation covers the plan it was given; where the stack re-presents a plan because the candidates changed, it is a new plan asking for a new confirmation.

N12 already forbids the app proposing a set, pre-selecting, or acting on a threshold (N12-R4, N12-R9).

Playback statistics: whose, and how sure

What was watched, by whom and how far through is the household’s most personal record (H8). Whose statistics a person may see is the core’s answer for the identity signed in, and the app holds no rule of its own about it (N3-R2). A device that reports no progress gives unknown, not watched and not unwatched. Somebody with no recorded playback is said to have none rather than shown an empty chart, and a history that begins where backfill could not reach says where it begins (N11-R1).

What somebody watched is never in a notification and never drawn behind the lock.

Metrics: the pipeline’s proof is shown here; the dashboards are not

The metrics pipeline proves three things: each target is up in the collector, each service’s series are present in the store, and each curated dashboard’s panels resolve against real data (K1). The app shows each, per target and per dashboard, with the remedy for each failure. A running exporter is not a healthy target, a service with no exporter is unmonitored rather than covered, and a collector scraping without retaining is not a healthy pipeline. The retention bound and the projected growth of the store are shown. Regenerating targets and dashboards, and running the proof, are offered from here.

The dashboards themselves are not drawn in the app. They are a bundled renderer’s own screens, generated as files from the stack’s definition (K1-R4), and lemonfiber’s other surfaces do not draw them either. Drawing them here would mean the app holding a second copy of every provisioned query, which drifts from the first the day a dashboard is regenerated, and the app renders the core’s answers rather than keeping copies of its own (N1-R1). Where the stack gives the renderer’s address, the app offers it as that address; it never builds one.

States

Edge cases

• A cross-seed pass where one indexer drops mid-way. That indexer is failed for the pass; the matches the others found stand.
• A filter authored to match everything. Still capped; the cap is shown holding rather than the filter shown as healthy.
• A catalogue source unreachable during a sync. Nothing was applied, and the instance is shown untouched rather than partly synced.
• A cleanup item removed by hand before the plan ran. Already gone, not a failed deletion.
• A member with no recorded playback. Said, not charted as zero.
• A target inside the VPN namespace. Its state is the collector’s, across the path it is scraped on; it is not shown down for being inside the tunnel.

Requirements

Notes

The contract carries none of this. Every feature on this page is planned, and no envelope the stack publishes is about any of them: nothing names a cross-seed match, an announce channel, a profile sync, a cleanup candidate, a playback or a metric. Every row here waits on the contract, and N1-R17 stops the work until it carries what the row needs. Nearby envelopes are not substitutes:

N26-R19 is the one exclusion on this page, and it is written as one because parity is otherwise the rule (N1-R2). The pipeline’s proof is offered here in full; the dashboards are a bundled service’s own screens, which no lemonfiber surface draws.

Related

• N1 — parity, and rendering the core’s answers
• N12 — removal, and what the app may not propose
• N16 — queue self-healing
• N11 — a record’s horizon
• N3 — what a member sees
• N4 — notifications and the lock
• N24 — choosing a quality preset
• H1, H2, H3, H6, H8 — the glue
• K1 — metrics and dashboards