79 improve non ciritcal gaps sections

.env.example.minimal

@@ -82,7 +82,6 @@ MAX_VISITS_PER_NODE=1
 # Extra bounded repair calls per facts unit, driven by machine-found
 # violations. Fires even at MAX_VISITS_PER_NODE=1, so a facts unit costs up to
 # two provider calls by default. Set 0 for exactly one call per unit.
-FACTS_LLM_REPAIR_VISITS=1
 
 # Concurrent content-unit workers. Bounded in practice by LLM_MAX_INFLIGHT
 # above; raising this alone will not make a rate-limited run faster.
@@ -148,7 +147,7 @@ AGG_EMBEDDING_MODEL=sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
 #EMBEDDING_DIMENSION=384
 
 # How much schema each unit is shown. The two knobs that actually move recall.
-#VECTOR_STORE_TOP_K=20
+#VECTOR_STORE_TOP_K=40
 #VECTOR_STORE_INDUCED_SUBGRAPH_MAX_TOTAL_TRIPLES=1200
 
 # Drop and rebuild the partition on startup. Required after any change to the

.github/workflows/pre-commit.yml

@@ -15,19 +15,24 @@ jobs:
             with:
                 version: "0.10.12"
                 enable-cache: true
-        # `ty check ontocast test` is a whole-program check, so every extra
-        # whose code it walks has to be importable here: `server` (FastAPI app
-        # + CLI), `documents` (docling-core types), `qdrant` and `graph`.
-        # Syncing `dev` alone -- correct before the dependency tiering moved
-        # those packages out of the base install -- made ty fail on ~40
-        # unresolved imports.
-        #
-        # This deliberately stops short of `--all-extras` (8.8 GB vs 633 MB;
-        # torch, docling/easyocr, lancedb, pygraphviz) because the heavy tier
-        # is instead whitelisted in `[tool.ty.analysis] allowed-unresolved-imports`
-        # in pyproject.toml. A new unresolved-import failure here means an
-        # extra crossed that line: either add it below, or whitelist it there.
-        -   name: Install dependencies (extras that ty type-checks)
-            run: uv sync --extra dev --extra server --extra documents --extra qdrant --extra graph --no-install-project
+        # `setup-uv`'s own cache covers uv's package cache, not the hook
+        # environments, which live in ~/.cache/pre-commit. Without this step
+        # every run rebuilt them from scratch and re-resolved each hook's
+        # dependencies, so an unchanged repository could go red on a transitive
+        # release. Keying on the config hash means a hook or pin change rebuilds
+        # and everything else reuses.
+        -   name: Cache hook environments
+            uses: actions/cache@v4
+            with:
+                path: ~/.cache/pre-commit
+                key: pre-commit-${{ hashFiles('.pre-commit-config.yaml') }}
+        # This job is lint and formatting only, so it needs nothing but the
+        # `dev` extra. Type checking runs in `tests.yml`, whose `unit` job
+        # already installs every extra -- a whole-program check gets better
+        # answers where the whole program is actually importable.
+        -   name: Install dependencies
+            run: uv sync --extra dev --no-install-project
         -   name: Run pre-commit
             run: uv run pre-commit run --all-files
+            env:
+                SKIP: ty

.github/workflows/tests.yml

@@ -97,6 +97,16 @@ jobs:
                 enable-cache: true
         -   name: Install dependencies
             run: uv sync --all-extras
+        # `ty check ontocast test` is a whole-program check, so it belongs where
+        # the whole program is installed. Under a partial install the optional
+        # tier has to be whitelisted as unresolved, which types those modules as
+        # `Any` and silently stops ty checking any call into them; here every
+        # extra is already present for the test run, so the check costs one step
+        # and no extra install. One matrix leg is enough -- the source is the
+        # same on both, and 3.13 only runs on main.
+        -   name: Type check
+            if: matrix.python-version == '3.12'
+            run: uv run ty check ontocast test
         # No .env exists in a clean checkout; pytest-dotenv no-ops on a missing
         # file, and integration tests self-skip when their service is
         # unreachable, so this runs offline.

.pre-commit-config.yaml

@@ -37,15 +37,25 @@ repos:
   # uv.lock is excluded: it is generated by `uv lock` in uv's own format, so
   # reformatting it here made every dependency change fail CI and then be
   # rewritten by the next uv run.
--   repo: https://github.com/macisamuele/language-formatters-pre-commit-hooks
-    rev: v2.16.0
-    hooks:
-    -   id: pretty-format-toml
-        args: [--autofix]
-        exclude: ^uv\.lock$
+  #
+  # `toml-sort` is the only TOML formatter here, and that is the point.
+  # `pretty-format-toml` from language-formatters-pre-commit-hooks used to run
+  # alongside it, but that hook is a thin wrapper around toml-sort -- so the
+  # suite was running one formatter twice, through two independently resolved
+  # copies of it. They drifted apart: toml-sort 0.25.0 swapped its engine from
+  # tomlkit to tomlrt, which writes inline tables as `{ k = v }` where tomlkit
+  # writes `{k = v}`. With the wrapper on a fresh resolve and this hook pinned
+  # by `rev`, each one reverted the other's spacing on the `authors` table,
+  # every run, forever. A single formatter cannot disagree with itself.
+  #
+  # tomlkit is pinned because `rev` does not pin it: pre-commit resolves a
+  # hook's dependencies when it builds the environment, so a months-old cached
+  # environment and a fresh CI one can format differently from identical config
+  # -- which is how this passed locally while failing in CI. Bump deliberately.
 -   repo: https://github.com/pappasam/toml-sort
     rev: v0.24.3
     hooks:
     -   id: toml-sort
         args: [-ia]
         exclude: ^uv\.lock$
+        additional_dependencies: [tomlkit==0.15.1]