Stop Think from managing git-warp cache

ADVANCED_GUIDE.md

@@ -6,10 +6,10 @@ For orientation and the productive-fast path, use the [GUIDE.md](./GUIDE.md).
 
 ## Graph-Native Doctrine
 
-Think uses a "window-based" read model to ensure the TUI remains responsive even as your archive grows to thousands of entries.
+Think uses bounded product read models to ensure recall and browse remain responsive as your archive grows.
 
-- **No Whole-Graph Materialization**: Commands like `browse` and `remember` use `git-warp` read handles to query specific worldline slices.
-- **Checkpoint-Backed Reuse**: The browse TUI uses checkpoints to avoid redundant traversal of chronological neighbors.
+- **No Whole-Graph Materialization**: Default `browse`, `recent`, and `remember` paths read Think-owned latest/recent read-model facts. Recall uses self-contained fast capture records so it does not hydrate capture nodes just to score ambient project matches.
+- **Runtime-Owned Acceleration**: `git-warp` owns checkpoints, state-cache, snapshots, and materialization; Think stores domain facts such as "latest capture" and "recent capture records."
 - **Immutable Raw Layer**: The core Git repository is the system of record. Raw thoughts are never modified after capture.
 
 ## Derivation Pipeline

CHANGELOG.md

@@ -12,6 +12,18 @@ Release discipline:
 
 ## Unreleased
 
+- added bounded Think read-model facts for latest/recent captures plus
+  self-contained fast capture records so default `--remember`, `--recent`, and
+  browse bootstrap avoid scanning `entry:*`
+- added ratchet tests proving default ambient recall and explicit recall avoid
+  graph queries, keyword wildcard scans, and capture-node hydration when fast
+  capture records are available
+- added a Hexagonal Boundary lint ratchet so product code cannot increase
+  direct substrate-layout or `git-warp` runtime leakage
+- removed Think-managed `git-warp` cache/checkpoint reads, doctor checks, and
+  ref deletion from product runtime paths
+- upgraded `@git-stunts/git-warp` to 18.2.1
+
 ## [0.7.2] - 2026-06-23
 
 - added feature proposal packets for History product boundaries, Browse memory workbench, followthrough job queue, agent-native memory API, and evidence-bound enrichment

docs/GLOSSARY.md

@@ -106,7 +106,9 @@ Git handles the repository substrate. `git-warp` provides the graph layer on top
 
 The primary application handle exposed by `git-warp`.
 
-In `think`, product read paths begin from `WarpApp`.
+`think` does not use `WarpApp` as a product read facade. Product reads open
+public `git-warp` worldlines and let `git-warp` own internal checkpoint and
+cache state.
 
 ### worldline
 
@@ -118,25 +120,27 @@ The normal product read handle in `git-warp`.
 
 A narrower read aperture derived from a `worldline`.
 
-In `think`, product reads use `WarpApp -> worldline() -> observer(...)` where that helps keep read scope honest.
+In `think`, product reads open a public `git-warp` worldline and read through
+that worldline. Think does not open `WarpApp` as a product read facade.
 
 ### core
 
 The whole-state inspection and admin-style `git-warp` surface.
 
-In `think`, `core()` is reserved for:
+In `think`, `core()` is reserved for explicit operator tooling such as:
 
 - migration
 - admin-style full-state inspection
-- narrow content-attachment reads not available on `worldline` / `observer`
 
 It should not be the default product read path.
 
 ### checkpoint
 
-`git-warp` recovery state that makes reopening read handles much faster.
+`git-warp` recovery state that can accelerate reopened read handles.
 
-Enabling checkpoints was one of the major browse-performance fixes in M4.
+Think product code does not inspect, delete, or repair checkpoint/state-cache
+refs. Routine reads let `git-warp` own cache decisions; the only Think code that
+names legacy checkpoint refs is the explicit v17 repair lane for old minds.
 
 ### history model version
 

docs/INFRASTRUCTURE_DOCTRINE.md

@@ -60,6 +60,19 @@ Examples of capabilities that require ports:
 - CLI, MCP, macOS, and browser I/O.
 - Encoding, decoding, and content storage.
 
+Product code must not know substrate layout. References to concrete
+storage engines, `git-warp` runtime classes, WARP ref paths,
+checkpoint/state-cache mechanics, materialization strategies, or repair
+commands belong only in explicit adapter, repair, or benchmark boundaries.
+Think diagnostics must not inspect or mutate `git-warp` cache refs; operators
+should use `git-warp` diagnostics for that substrate. CLI, MCP, Browse,
+Remember, Stats, and domain modules consume product ports and runtime-backed
+domain facts, not storage internals.
+
+This boundary is enforced by `npm run hexagonal-boundary:ratchet`, which
+runs from `npm run lint`. Existing leaks are debt in the ratchet baseline;
+new code must not increase them.
+
 ### 2. Dependency Injection
 
 Dependencies are injected through constructors or semantically named

docs/audit/hexagonal-boundary-ratchet-baseline.json

@@ -0,0 +1,38 @@
+{
+  "allowedBoundaryFiles": [
+    "scripts/hexagonal-boundary-ratchet.mjs",
+    "scripts/repair-v17-mind.mjs",
+    "src/browse-benchmark.js",
+    "src/browse/adapters/git-warp-worker.js",
+    "src/browse/adapters/git-warp.js",
+    "src/cli/commands/doctor.js",
+    "src/doctor.js",
+    "src/history/git-warp-read.js"
+  ],
+  "generatedFrom": "scripts/hexagonal-boundary-ratchet.mjs",
+  "leaks": {
+    "byFile": {
+      "src/git.js": 1,
+      "src/store/annotate.js": 4,
+      "src/store/capture.js": 6,
+      "src/store/derivation.js": 3,
+      "src/store/enrichment/runner.js": 2,
+      "src/store/migrations.js": 6,
+      "src/store/reflect.js": 6,
+      "src/store/runtime.js": 11
+    },
+    "byTerm": {
+      "git-warp-package": 1,
+      "git-warp-runtime-symbol": 37,
+      "warp-ref-layout": 1
+    },
+    "total": 39
+  },
+  "substrateTermIds": [
+    "git-warp-package",
+    "git-warp-runtime-symbol",
+    "checkpoint-mechanic",
+    "warp-ref-layout"
+  ],
+  "version": 1
+}

docs/audit/runtime-truth-ratchet-baseline.json

@@ -8,8 +8,8 @@
   "strictLimits": {
     "byCategory": {
       "benchmark": 5,
-      "source": 150,
-      "test": 45
+      "source": 142,
+      "test": 34
     },
     "byFile": {
       "benchmarks/browse-bootstrap.js": {
@@ -158,12 +158,10 @@
       },
       "src/cli/commands/capture.js": {
         "byRule": {
-          "complexity": 1,
-          "max-lines-per-function": 1,
-          "max-statements": 1
+          "max-lines-per-function": 1
         },
         "category": "source",
-        "total": 3
+        "total": 1
       },
       "src/cli/commands/read.js": {
         "byRule": {
@@ -215,10 +213,10 @@
       },
       "src/mcp/service.js": {
         "byRule": {
-          "max-lines-per-function": 2
+          "max-lines-per-function": 1
         },
         "category": "source",
-        "total": 2
+        "total": 1
       },
       "src/splash-shader.js": {
         "byRule": {
@@ -231,13 +229,6 @@
         "category": "source",
         "total": 17
       },
-      "src/store/checkpoint-product-read.js": {
-        "byRule": {
-          "complexity": 1
-        },
-        "category": "source",
-        "total": 1
-      },
       "src/store/derivation.js": {
         "byRule": {
           "complexity": 2,
@@ -314,19 +305,10 @@
       },
       "src/store/runtime.js": {
         "byRule": {
-          "complexity": 7,
-          "max-lines-per-function": 1
+          "complexity": 4
         },
         "category": "source",
-        "total": 8
-      },
-      "test/acceptance/graph-migration.test.js": {
-        "byRule": {
-          "complexity": 2,
-          "max-lines-per-function": 7
-        },
-        "category": "test",
-        "total": 9
+        "total": 4
       },
       "test/acceptance/json-output.test.js": {
         "byRule": {
@@ -382,14 +364,6 @@
         "category": "test",
         "total": 2
       },
-      "test/ports/checkpoint-read.test.js": {
-        "byRule": {
-          "max-lines-per-function": 1,
-          "max-statements": 1
-        },
-        "category": "test",
-        "total": 2
-      },
       "test/ports/docs-consistency.test.js": {
         "byRule": {
           "max-lines-per-function": 1
@@ -407,14 +381,14 @@
       }
     },
     "byRule": {
-      "complexity": 61,
+      "complexity": 54,
       "max-depth": 11,
       "max-lines": 1,
-      "max-lines-per-function": 96,
+      "max-lines-per-function": 86,
       "max-params": 3,
-      "max-statements": 28
+      "max-statements": 26
     },
-    "total": 200
+    "total": 181
   },
   "strictRuleIds": [
     "complexity",

docs/design/README.md

@@ -68,7 +68,7 @@ Completed milestone and slice notes that still matter as historical context, but
 - [`../retrospectives/m4-session-traversal.md`](../retrospectives/m4-session-traversal.md): closeout for the M4 slice that makes same-session movement a first-class browse behavior for both humans and agents.
 - [`../retrospectives/m4-graph-migration-gating.md`](../retrospectives/m4-graph-migration-gating.md): closeout for the slice that makes migration explicit for graph-native commands while keeping raw capture migration-safe.
 - [`../retrospectives/m4-v3-read-edge-substrate.md`](../retrospectives/m4-v3-read-edge-substrate.md): closeout for the narrower sub-slice that lands `v3` graph-native read edges without yet claiming the browse bootstrap performance win.
-- [`../retrospectives/m4-graph-native-browse-read-refactor.md`](../retrospectives/m4-graph-native-browse-read-refactor.md): closeout for the slice that moves product reads onto `WarpApp -> worldline() -> observer(...)`, enables checkpoint-backed browse startup, and records the official `AFTER` benchmark.
+- [`../retrospectives/m4-graph-native-browse-read-refactor.md`](../retrospectives/m4-graph-native-browse-read-refactor.md): historical closeout for the slice that first moved browse reads behind a runtime adapter and recorded the official `AFTER` benchmark.
 - [`../retrospectives/m4-graph-migration-progress-ux.md`](../retrospectives/m4-graph-migration-progress-ux.md): closeout for the slice that makes the human interactive upgrade moment visibly in progress while keeping agent and capture behavior unchanged.
 - [`../retrospectives/m4-remember-enhancements.md`](../retrospectives/m4-remember-enhancements.md): closeout for the slice that adds bounded and brief remember recall without changing the underlying ranking model.
 - [`../retrospectives/m4-prompt-telemetry-read-surface.md`](../retrospectives/m4-prompt-telemetry-read-surface.md): closeout for the slice that turns recorded prompt telemetry into a factual CLI / JSON read surface without drifting into dashboards.

docs/method/backlog/bad-code/CORE_boundary-codec-cutover.md

@@ -13,16 +13,15 @@ payload encoding directly in domain-adjacent modules.
 
 Examples include `Buffer.from()` in `src/store/content.js`,
 `parseJsonArray()` in `src/store/model.js`, JSON-string graph
-properties in derivation flows, and content byte decoding in checkpoint
+properties in derivation flows, and content byte decoding in runtime
 read models.
 
 ## Acceptance Criteria
 
 - Introduce named codec ports for text content, graph property payloads,
-  and checkpoint content reads.
+  and runtime content reads.
 - Keep `Buffer`, `TextEncoder`, `TextDecoder`, JSON parsing, and future
   CBOR work in adapters or codec modules only.
 - Core store workflows receive validated domain objects, not raw decoded
   transport shapes.
-- Existing content and checkpoint-read tests pass through the new codec
-  ports.
+- Existing content and runtime-read tests pass through the new codec ports.

docs/method/backlog/bad-code/CORE_git-warp-dependency-truth.md

@@ -5,26 +5,23 @@ blocked_by:
   - CORE_repair-v17-git-warp-minds
 ---
 
-# git-warp dependency truth is split between package metadata and local v17 links
+# git-warp dependency truth must stay anchored to package metadata
 
-Think currently declares `@git-stunts/git-warp` as `15.0.0`, while local
-development can resolve to a linked `17.0.0` checkout. That makes
-`npm ls @git-stunts/git-warp` fail with `ELSPROBLEMS` and leaves runtime
-compatibility depending on local workspace state rather than package truth.
-
-The checkpoint read path now includes a public-reader compatibility bridge for
-`createStateReader` vs `createStateReaderV5`. That bridge is acceptable as a
-short-term guard, but it should not become permanent dependency sludge.
+Think runtime compatibility must be proven against the published
+`@git-stunts/git-warp` package declared in `package.json`, not against a linked
+local checkout or private package internals. The current product runtime targets
+the public worldline API in `@git-stunts/git-warp@18.2.1`; legacy checkpoint
+repair remains quarantined in the v17 repair lane.
 
 ## Acceptance Criteria
 
 - `npm ls @git-stunts/git-warp` exits cleanly in a normal checkout.
 - `package.json` and `package-lock.json` match the intended git-warp version.
 - The intended version is published or resolved through an explicit,
   documented local/workspace dependency path.
-- Checkpoint read tests pass from a clean install, not only from a local
+- Product read/write tests pass from a clean install, not only from a local
   linked git-warp checkout.
 - The archived v17 repair acceptance fixture runs its full repair assertion in
   clean CI instead of skipping when the v17 migration package is unavailable.
-- The state-reader compatibility bridge is either documented as intentional
-  version support or removed after the dependency cutover.
+- No production module imports private `node_modules/@git-stunts/git-warp/src`
+  paths or git-warp cache/checkpoint compatibility helpers.

docs/method/backlog/bad-code/CORE_hexagonal-store-boundary.md

@@ -17,6 +17,9 @@ on injected ports and runtime domain objects, not on concrete host APIs.
 
 ## Acceptance Criteria
 
+- Add a lint ratchet that counts direct substrate-layout and `git-warp`
+  runtime leakage outside explicit adapter, diagnostic, repair, or
+  benchmark boundaries.
 - Define explicit ports for graph persistence, content storage, clocks,
   random IDs, host metadata, prompt metrics storage, and ambient project
   context.
@@ -28,3 +31,10 @@ on injected ports and runtime domain objects, not on concrete host APIs.
   `@git-stunts/*` concrete adapters directly.
 - Add at least one browser-like unit test that runs core store logic
   without Node filesystem, process, or Git globals.
+
+## Progress
+
+- [x] Hexagonal Boundary ratchet added to `npm run lint`.
+- [x] Product runtime paths stopped inspecting, deleting, or reading through
+      `git-warp` checkpoint/state-cache refs.
+- [ ] Existing substrate leaks moved behind product ports or adapters.

docs/method/backlog/bad-code/CORE_large-mind-read-timeouts.md

@@ -7,6 +7,10 @@ read-heavy paths such as `recent`, `stats`, and `doctor`. The Claude mind at
 `~/.think/claude` was observed with roughly 43k loose Git objects and read
 commands taking 14-21 seconds even after the raw capture path itself was fast.
 
+Default recall also regressed when `remember` asked broad `entry:*` and
+`keyword:*` questions, then sorted and sliced after the runtime had already
+answered the broad query.
+
 ## Why It Matters
 
 Capture can return after local raw save, but agents still need reliable re-entry
@@ -15,10 +19,21 @@ unavailable even when the underlying mind is intact.
 
 ## Acceptance Criteria
 
+- [x] Route default `remember`, `recent`, and browse bootstrap through bounded
+      Think read-model facts instead of broad capture wildcard scans.
+- [x] Add port-level ratchets proving default ambient and explicit remember do
+      not call `read.view.query()`.
+- [x] Store self-contained fast capture records so default `remember` can score
+      ambient project matches without hydrating capture nodes.
 - [ ] Add a deterministic large-mind fixture or synthetic benchmark for MCP read
       timeout budgets.
+- [ ] Add an explicit read-model backfill/repair command for existing minds so
+      old captures can enter the bounded indexes without default recall scanning
+      from genesis.
 - [ ] Establish target budgets for `recent`, `stats`, `doctor`, and `remember`
-      against repaired checkpoint-backed minds.
+      against large repaired minds. Current `codex-think --remember --json`
+      smoke exits under a 20s timeout on `~/.think/codex`, but still spends
+      roughly 13-14s in one git-warp optic property read.
 - [ ] Document and automate safe maintenance for high-loose-object minds.
-- [ ] Prefer checkpoint-backed bounded reads where whole-graph observer startup is
-      not required.
+- [ ] Prefer public worldline/optic bounded reads where broad transitional
+      queries are not required.

docs/method/backlog/cool-ideas/SURFACE_document-window-based-read-model.md

@@ -1,8 +1,8 @@
 # Document window-based read model
 
-Add a "Window-Based Navigation" section to ADVANCED_GUIDE.md
-explaining how git-warp read handles prevent whole-graph
-materialization, how the browse TUI loads neighbors lazily, and
-how checkpoint-backed reuse keeps navigation fast.
+Add a "Window-Based Navigation" section to GUIDE.md or the design docs
+explaining how public git-warp worldline/optic reads prevent unnecessary
+whole-graph materialization, how the browse TUI loads neighbors lazily, and
+where broad transitional queries still need caveats.
 
 Source: documentation-quality audit 2026-04-11 §2.3.