perf: build declaration inlays from model facts

docs/05-type-model.md

@@ -47,6 +47,13 @@ returned to a caller holds no type-expression node, unless it was parsed with
 each would only add memory, 2.4 MB on the `validate` workload's `service`
 configuration. Read only `value` and `position`, which both forms have.
 
+`BaseShape.type_written` records authored type syntax during declaration decoding,
+before resolution or unwrap, and survives `clone`/`clone_detached`. Scalar type
+names and multiple-inheritance sequences are explicit, as are accepted `type:`
+and `schema:` fields even when their value is null. A bare empty/null declaration
+or a mapping with only inferred facets has no authored type. This fact is distinct
+from the resolved kind or the presence of bound expression references.
+
 ## 2. Kinds and facets
 
 | Kind | Kind-specific facets |

docs/13-public-api.md

@@ -162,6 +162,10 @@ Consumers must honor these contracts:
    partial: the expression's `value` and `position`, not a YAML `Node`. With
    `retain_source` it stays the `Node`. Read only `value` and `position`
    (docs/05 § 1).
+6. `BaseShape.type_written` records whether a declaration supplied type syntax,
+   independently of its resolved kind or expression references. Cloning preserves
+   it; bare empty/null declarations are inferred, while an explicit `type:` or
+   `schema:` field is authored even with a null value (docs/05 § 1).
 
 Parsing, `build_graph`, `build_occurrences`, `Linter` runs, `to_openapi`, and
 the `lsp` verb raise the garbage collector's full-collection threshold while

docs/15-implementation-plan.md

@@ -22,7 +22,13 @@ XML Schema external types are unsupported ([01](01-scope-and-coverage.md) § 3).
 **Service baseline acceptance and recovery.** The simpler source-cache branch
 passes the master-controlled rebuild and mixed-request checks after sharing its
 query source owner. The user accepted its documented first-use trade-off;
-CI verification and integration are pending.
+all CI checks passed and the baseline is merged. The first isolated recovery
+experiment implements model-backed declaration inlays against that accepted
+control. Its correctness, scaling and A/B results are recorded in the
+[inlay recovery report](reports/2026-10-10/service-model-inlays.md). The user
+accepted its snapshot-first mixed allocation-peak trade-off. Remaining recovery
+candidates are shared semantic/hierarchy indices and outline caching, cursor-local
+source lookup, and accurate authored section ranges, each measured independently.
 The record-backed representation replacement remains parked; its recurring
 rebuild and first-use costs are recorded in the
 [service cost review](reports/2026-10-10/service-authoring-cost-review.md).
@@ -75,8 +81,9 @@ the media-type fixes and the type-walk work landed together:
   owner removes the old duplicate hover/structural tree. The accepted baseline's
   remaining first-use cost and post-query allocation trade-off are recorded in the
   [integration report](reports/2026-10-10/service-baseline-integration.md).
-  Model-backed declaration inlays and cursor-local source lookup are the next
-  separately measured recovery candidates.
+  Model-backed declaration inlays remove that first-use work from hints alone;
+  their accepted request-order peak trade-off is recorded in the inlay recovery
+  report. Cursor-local source lookup remains a separately measured candidate.
 
 ## 3. Potential future work
 

docs/21-language-service.md

@@ -56,13 +56,14 @@ the TCK's 1011 files.
 `parser:` limits. It keeps the YAML trees too, `retain_source=True`, only when
 an enabled lint rule reads them; no rule in the default set does. Decoders
 record the small syntax facts that `deprecated-schemas` reads (docs/18 § 6).
-The first query that needs a file's tree composes it on demand. Hover, inlays,
+The first query that needs a file's tree composes it on demand. Hover,
 folding and selection share the workspace's source owner when their input and
 composition policy agree (§ 4). Parsing does not populate that query cache or
 retain all producer trees. Snapshot-local hover indices keep requested trees;
 snapshots borrow the workspace's cache weakly rather than keep its other files
 alive. Without a live owner, a held snapshot can compose its own retained text.
-It records the set of files it read: every retained text, every
+Inlays read model facts and typed data without requesting a tree (§ 4.4).
+A snapshot records the set of files it read: every retained text, every
 fragment, and every include it tried, found or not. It is built when a query
 first asks for it and kept until one of those files changes. A file that
 appears, a buffer opened or a file created, also drops every snapshot that
@@ -350,6 +351,18 @@ returns hints only within the requested source range. It reuses hover's
 authored source sites and typed `DataNode` token index, and never resolves a
 name or repeats type matching in the adapter.
 
+The decoder's `type_written` flag distinguishes explicit type syntax from
+inference. Declaration hints read per-file authored model subjects and expression
+spans, without composing source or populating source grammar/built-in indices.
+Inferred hints need a real single-line key span; a fragment-root body span is not
+an authored declaration key. Reference hints still use the authored expression
+span in that fragment.
+
+Expanded YAML aliases keep their anchor's child spans. Declaration hints omit
+those borrowed children, including nested object and array members, while keeping
+any declaration hints authored at the anchor itself. This check reads model
+placement and owned declaration edges; it does not build the source grammar.
+
 A declaration without an authored type can show its inferred type, such as
 `[object]`. Explicit built-in types and multiple-inheritance lists are not
 repeated. Named type references show the underlying effective type

docs/reports/2026-10-10/service-baseline-integration.md

@@ -145,5 +145,43 @@ against that accepted baseline, without accepting record capture.
 
 ## 3. Selective recovery
 
-Pending acceptance of the production baseline. The first planned experiment is
-model-backed declaration inlays and `type_written`, without record capture.
+All production-baseline CI checks passed, including Linux, Windows, pure-Python
+YAML, TCK and benchmarks. [PR #2](https://github.com/deiteris/FastRAML/pull/2)
+merged at `bb6941fce593082a4b38a33f4b346dc35268a8d9`. This exact revision is the
+comparison control for the first recovery experiment.
+
+### 3.1 Model-backed declaration inlays
+
+Branch: `perf/service-model-inlays`, based on the accepted baseline. Objective:
+remove the whole-file grammar/source construction that inlays currently request
+just to identify authored types and declaration sites. This does not port shared
+semantic indices, accurate outline sites or the record-backed representation.
+
+The decoder records whether a declaration supplied its type; clones preserve the
+fact. A bare empty/null declaration is inferred, while an authored `type:` key
+(including its null value) is explicit syntax. The existing hover subject/sites
+index supplies declarations, with per-URI subject references to avoid scanning
+the whole model for each queried file. Inferred labels require an authored
+single-line key; fragment-root body spans must not become invented declaration
+keys. Named-reference hints continue to use the model's type-expression span.
+
+Expected work: one boolean slot/assignment per shape and clone, one per-URI
+subject reference per distinct authored site when hover is first populated, one
+hint construction per declaration in the queried file/snapshot, then range
+lookup. Inlays should make zero source compositions or grammar/built-in-token
+index populations. Typed-value navigation remains lazy and shared as before.
+
+Acceptance compares `inlays`, both mixed request orders, ordinary parse/validate
+and routine `large/service` against `bb6941f`. A reach test must forbid source
+construction during both timed and allocation inlay passes. Behavior cases cover
+explicit versus inferred syntax, multiple inheritance, cloning, included types
+and template materializations. Preserve the remaining gates and record the
+allocation cost of the new model flag, even when time stays within noise.
+
+The isolated port and measurements are now recorded in the
+[inlay recovery report](service-model-inlays.md). Cold inlay time falls 27.4%,
+peak 34.8% and retained allocation 30.2%; recurring rebuild time stays within
+noise. A 10.0% snapshot-first mixed peak increase is attributed to query-order
+allocation overlap, with final retention effectively unchanged. After the
+shared-snapshot/order explanation, the user explicitly accepted that new trade-off
+and instructed integration on 2026-10-10; normal CI verification gates the merge.