Adopt externally-managed OpenAI-compatible endpoints by base URL

[vonargo]

Adopt externally-managed OpenAI-compatible endpoints by base URL

Part of #24 (minimal slice). This adds a way to adopt a server that already speaks the
OpenAI HTTP API — vLLM, llama.cpp server, TGI, or anything similar — by declaring its
base URL, without PAIR needing to know which software sits behind it. vLLM is the test
fixture, not the feature.

What changed

• nvpair-manual-nodes (0.11.1 → 0.12.0) — node/add accepts exactly one of
  address (historical form: host probed on the default engine ports) or
  openai_base_url (http only in this release). Declared endpoints are probed at
  their own URL: GET {base}/models doubles as liveness + model list
  (openai_up / openai_models), plus best-effort node-info on the URL's host.
  The default engine-port legs are not probed for this entry type. The service now
  owns its entry list in the app data directory (manual-nodes.json, atomic
  tmp+rename) and restores it at startup, so manual nodes survive a restart.
• nvpair-ui-broker (0.40.2 → 0.41.0) — bridges declared endpoints into
  lmstudio-proxy with their path prefix; attributes openai_models to the new
  "openai" engine in modelsByEngine; translates node/remove by UUID back to
  the manual store key for the endpoint leg. An endpoint keeps its manual key
  (the declared URL's identity) and never adopts the host's learned node-info
  UUID: an endpoint on the same box as PAIR — or two endpoints on one host —
  must not fold into that machine's own node, where they would clobber each
  other on every probe.
• lmstudio-proxy (0.16.2 → 0.17.0) — manual nodes may carry base_path
  (e.g. /v1). The /v1 prefix is the proxy's own API root, so inbound
  /v1/... paths are joined onto the endpoint's root for forwarding and model-list
  fanout; nodes without a base path forward verbatim as before. Workloads served by
  a base-prefixed endpoint are labeled engine: "openai" (including on failover).
• nvpair-tui (0.7.2 → 0.9.0) — the add prompt accepts a bare host or
  http://host:port/v1; the manual table gains an OPENAI reachability column.
  Also fixes a pre-existing table overflow found while testing: the upstream
  table styling rendered rows wider than the widths the existing tab layouts
  budgeted, and the table's viewport truncates rows at the terminal edge. The
  problem was mostly hidden because the truncated region often contained
  trailing whitespace; the added OpenAI column moved real content into the
  clip zone and made the clipping visible. Cells now render at their exact
  budgeted widths.
• desktop — the Add-node dialog gains an "OpenAI endpoint" row (input + Add)
  that invokes the new nodes:add-endpoint ws channel, which relays
  node/add { openai_base_url } and returns inline errors. One-line mapping of
  workload engine openai onto the existing lm-studio display path so
  external-endpoint jobs are not dropped by the closed engine union.
• stability (found while dogfooding) — same-box endpoints were
  re-bridged into the proxies (remove + add) on every 10s probe: the
  prober's change diff counted msSince (the GPU sample's age, which
  advances every probe by construction), so every node-info-up manual
  node emitted a state change each cycle, and the broker re-issued the
  bridge unconditionally. msSince is now excluded from the diff
  (display-only — the store stays fresh for nodes/list), and the
  manual→proxy bridge skips the RPC when the exact intent (add payload
  or remove) was already applied to that proxy + engine + key slot.
• docs + versions — manual-nodes and lmstudio-proxy READMEs,
  docs/architecture.mdx manual-node paragraph, services/versions.json
  (product/installer 0.93.0).

Wire contract

• node/add param + status echo: openai_base_url
• manual-nodes status: openai_up, openai_host, openai_port, openai_base_path, openai_models
• proxy Node / node/add-manual: base_path
• workload engine / modelsByEngine key: "openai"
• desktop ws channel: nodes:add-endpoint { request: { url }; response: { ok, error? } }
• entry rule: exactly one of address | openai_base_url; http only (v1)
• identity rule: address entries re-key to the host's learned UUID (dedup with
  mDNS, as before); endpoint entries always keep their manual id

Verification

• Per-module Go suites green (fresh runs): nvpair-manual-nodes, nvpair-ui-broker,
  lmstudio-proxy (incl. a new real-binary e2e: models fanout + inference land on
  the endpoint's /v1-rooted paths without /v1/v1 doubling, workload labeled
  engine: "openai"), nvpair-tui; cross-process broker suite (bridge tests incl.
  the new endpoint bridge) green.
• Desktop: typecheck, test:unit (213 passing), service-contracts:check,
  dead-code:check, lint all clean.
• TUI width-sweep test: the manual table's last column survives the viewport
  truncation at widths {44, 60, 80, 120, 168} — regression guard for the
  overflow fix above.
• Churn regression guards: nvpair-manual-nodes (a probe that changes only
  the telemetry sample age emits no node/updated) and nvpair-ui-broker
  (bridge-intent equality: identical add payload / remove is equal, any
  routing-relevant difference is not).
• Service-level walkthrough (live vLLM at 127.0.0.1:8888):
  1. node/add {"openai_base_url":"http://127.0.0.1:8888/v1"} → initial unprobed
     echo with parsed host/port/base path; then node/discovered with
     openai_up: true and the endpoint's model inventory.
  2. lmstudio-proxy --port 18099 + node/add-manual {..., "base_path":"/v1"} →
     GET /v1/models on the proxy returns the endpoint's inventory.
  3. POST /v1/chat/completions on the proxy → routed to the endpoint (proxy log:
     node_id: "stub", target: "127.0.0.1:8888", 200), real completion returned.
  4. Kill the manager, restart, nodes/list → the entry comes back from
     manual-nodes.json (persisted + restored).

Known limitations (follow-ups, out of scope here)

• Endpoint nodes render with the "LM Studio" engine chip on desktop node cards;
  honest per-engine display is part of [Feature]: Adopt arbitrary OpenAI-compatible backends with endpoint manifests, model aliases, and deterministic capability routing #24's broader UI work.
• https endpoints are rejected (v1 is http only); no auth headers.
• Endpoints are not advertised to other peers' proxies (cluster-wide reachability
  of manual nodes is Make a node added by address a first-class peer (overlay networks such as Tailscale) #10 territory).
• The desktop's legacy manual-nodes.json store + replay remain in place
  (untouched; nothing on main wrote entries there). Removal is a follow-up cleanup.
• The desktop has an add path but no remove path for endpoint entries:
  node removal in the desktop is cluster-scoped, and endpoints are not cluster
  members. Remove one in the TUI Manual view (r on the selected row) or by
  editing the service store. A desktop-side affordance is follow-up UI work.
• No name field in the v1 add flows — the node ID defaults to
  manual:<host>:<port>; the service still accepts a name param for future UIs.
• Residual steady-state log lines (pre-existing, unchanged here): manual
  nodes still emit node/updated each probe cycle because that event also
  carries the node's telemetry to the scheduler (gating it would make a
  healthy endpoint age out as stale within the scheduler's 10s freshness
  window). The scheduler's periodic priority reconciliation and the
  proxies' local-backend refresh therefore still log on a regular cadence;
  the redundant bridge remove/add churn is gone. Tightening the scheduler's
  freshness window vs. probe cadence, and log verbosity in
  set-local-backend handling, are follow-ups.