1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
//! The one provider-row judgement every §9 gate calls ([`is_unknown_row`]).
//!
//! **This file was the `models:` half of the grammar** — yog's own table in
//! litany's `models.yaml`, one fact wide: the `context_window` the §5.1 #35
//! fullness figure divided by, authored by a Declare control and edited by
//! §9.5 typed rows. bl-9c8a deleted the table and everything over it. The
//! window is the provider's fact, and brazen now states it in band on every
//! `Usage` event litany already records per step (its model-discovery
//! §5.5), which is the number litany's own `window_percent` compaction trigger
//! divides by; a window yog declared on its own was a second representation of
//! that fact, one litany compacting the context could never see. The
//! figure reads the window off the step record (`budgets::context_window`), a
//! row that states none renders no figure, and the seat to state one is
//! upstream on brazen's provider row — where for Ollama it is already the
//! `num_ctx` in force.
/// **The** provider-row judgement: does `provider` name no row in brazen's
/// effective table? Every site that asks it asks it here — the §9.4 pick gate
/// ([`plan`](crate::model_pick::plan)), the §9.4 role marks
/// ([`role_fault`](crate::model_pick::role_fault)), and the §9.5 pane's provider
/// control over `providers.yaml` ([`crate::config_edit::form`]) — so the three
/// can never disagree. Each judges the LIVE pointer, `roles.<r>.provider`, which
/// is the whole of a role's binding (bl-35e2), against the wall of the workspace
/// that holds it. Two sites are gone and both for the same reason, a judgement
/// made where the answer was not: the birth gate (bl-c3a9, retired bl-00ee)
/// asked before a wall existed, and the §9.2 Apply gate (bl-53be, retired
/// bl-3ffa) asked about `models.<id>.provider`, a field nothing dispatches
/// through.
///
/// `providers` is `bz --list-providers`' answer — the workspace's **routing
/// table**: the rows its `config.toml` names plus the built-in presets a
/// credential resolves for (brazen bl-6feb). A built-in preset nobody named
/// and nobody signed into is NOT in it, and this judgement stays with the
/// table on purpose: binding a role to a row that cannot run is the defect
/// brazen just removed from its own default, and the §8.3 sign-in catalog is
/// where such a row is offered — so the refusal every caller words names the
/// two acts that put a row in the table, `/login` and a key in `/config
/// brazen`, never "add the row". An **empty** table is no answer rather than
/// an empty one — brazen could not be asked — so it judges nothing: no surface
/// may refuse on the strength of a question that went unanswered.