llm: routes that name a host, a backup and a window, and a profile that reads them ...

`llm.json` replaces `OLLAMA_BACKENDS_FILE`. It has targets (host and key
environment variable) and routes, where a route is an ordered chain of
`model @ target` entries, each with its own `context_window` and options. The
`default` route now starts on `glm-5.3-flash:cloud` and backs it with
`qwen3.6:35b-128k` on the home host, both at 131072 tokens, so a failover no
longer changes the budget the compressor and `num_ctx` are accounted against.
Embeddings get their own `embedding` route instead of a pinned host.

The engine classifies a failure (network, quota, auth, server, missing model),
cools that entry and moves the call down the chain; a background probe watches
only degraded targets and lets the next turn return to the primary once it is
healthy — it never interrupts a running turn. A profile joins a route with
`llm_route`, and the loader now actually reads it: `llm_route` and
`context_window` were declared on `AgentProfile` and written into every
`config.json` while the loader skipped both, the same silent no-op that
`is_admin_only` used to be, so every profile ran the legacy model list at the
global window. The writer emits them too, or an admin edit would erase the
route from disk.

The turn's resolved entry travels to the UI: the first chunk is stamped with
model and target, `model_info` carries `fallback` when a backup answered, and
the chat header shows a chip only then. `serving_primary` compares the target
as well as the model name, because a route may run one model on two hosts.
`/health` gains an `llm` block with the active entry, target states and missing
models.

`deploy/ollama/` keeps the local model's definition honest: MTP off (slower at
every context on this host), `presence_penalty 0` (the base model's 1.5 breaks
agentic tool calls), and `num_gpu 40` — Ollama's automatic fit claims a 128k
f16 slot fits and then dies mid-prefill with CUDA OOM; 40 of 42 layers leaves
~400 MiB and decodes at ~77 tok/s, 38 layers collapses to 48, and 39 OOMs.
q8_0 KV measured 37 tok/s against 70 on f16 here, so the cache stays f16.

startup, llm, profiles, todo: five more the prod logs gave up ...

The same 5-day sweep of prod's journal as 5e51f87, continued. Every one of
these failed silently: no error in the UI, just a line in the log or a hole
in the database.

complete()/embed() retry a connection error before falling back.
stream_complete has always given its first chunk two attempts, but the
non-streaming path made exactly one per (server, model), and with a single
configured server (ollama.com) one ReadTimeout raised "All backends
exhausted: ReadTimeout" outright. That is the whole story of 2026-10-08,
when 27 calls died that way in a day — 23 of them memory summarisation, the
rest planning. Attempts now come from settings.llm_complete_retries (2) with
llm_retry_backoff_sec (2.0) between them; a missing model is still not
retried (a 404 is an answer, not a hiccup), and a lone server is still never
blacklisted — blacklisting it would block the next request for _TTL.

Startup waits for Postgres. create_container() opens the pool eagerly and
the retry loop for the DDL tables sits below it, so a host that comes up
before docker takes the whole lifespan with it. On 2026-10-07 15:19:40 the
agent ran `sudo reboot` on itself at the user's request ("Перезагрузи
себя"); the host booted 15:20:03, navi started at 15:20:18, hit
ConnectionRefusedError on the pool and logged "Application startup failed.
Exiting." Connection errors now retry for ~30s; a genuinely broken config
(a missing DATABASE_URL) still fails fast.

The profile loader reports its tally. A profile dropped by an unreadable
config.json is skipped by design so it cannot take the server down, but on
that same 15:20 boot four of them — modeler_3d, navi_code, secretary,
server_admin — were dropped with `Extra data: line 125 column 1`, and the
only trace was one error line each: the UI simply showed four profiles
fewer.

todo recovers a dropped `op`. Of the 3385 todo calls stored on prod, 73
arrived with the discriminator flattened away ({"index": 1, "status":
"in_progress"}, {"action": "view"}, {"": "add", "tasks": [...]}) and got a
bare "Unknown op: None" — a wasted round trip for an intent the remaining
arguments state plainly. The error message now names the five ops and what
each needs.

Message.created_at is stamped at creation. It defaulted to None and only
some construction sites filled it, so 85% of stored rows had no timestamp at
all (every tool message, 14 168 of 14 168) and no query could slice history
by time or measure a pause. Loading a row passes the column through
explicitly, so NULL rows stay NULL instead of being restamped with "now".

Every new test was checked to fail with its fix reverted. Full suite green
(1466 passed, 1 skipped). No frontend or dependency change, so the deploy is
a pull and a restart.