Fix the open v0.22.0 review findings, and prove the profiler on real F# programs

.agents/skills/submit-pr/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: submit-pr
 description: Creates a pull request with a well-structured description after verifying CI passes. Use when the user asks to submit, create, or open a pull request.
-disable-model-invocation: true
+disable-model-invocation: false
 ---
 <!-- agent-pmo:a72c926 -->
 

.github/workflows/ci.yml

@@ -296,9 +296,9 @@ jobs:
   #
   # `always()` so it reports a real conclusion rather than inheriting a skip:
   # on a docs-only PR every upstream job skips and this passes; on a Dependabot
-  # PR the whole pipeline skips and this passes; any failure or cancellation
-  # anywhere upstream fails it, and so does a skipped build or test job when
-  # code changed. Either way the merge is blocked.
+  # PR the whole pipeline skips and this FAILS, since it ran nothing; any
+  # failure or cancellation anywhere upstream fails it, and so does a skipped
+  # build or test job when code changed. Either way the merge is blocked.
   ci:
     name: CI
     if: ${{ always() }}
@@ -338,6 +338,15 @@ jobs:
         run: |-
           echo "::error::Code changed but a build or test job was skipped. Check its if: condition."
           exit 1
+      # The whole pipeline skipped this Dependabot PR, so a pass would be a
+      # pass of nothing. Its bump reaches `main` only through the tested
+      # consolidation PR; the sweep retires this one. [DIST-CI-DEPENDABOT]
+      - name: Fail a Dependabot PR to main, which ran no job
+        if: ${{ github.actor == 'dependabot[bot]' }}
+        shell: bash
+        run: |-
+          echo "::error::Dependabot PRs run no CI. dependabot-automerge.yml sweeps this bump into dependabot-upgrades; it reaches main through the tested dependabot-upgrades -> main PR."
+          exit 1
       - name: Report
         shell: bash
         run: echo "All CI jobs passed or were skipped."

Makefile

@@ -40,7 +40,7 @@
 # sidecars ONCE and emitting that platform's VSIX and its standalone server
 # archive ([DIST-ARCHIVE]); VERSION is optional and defaults to 0.0.0. Then
 # `_print-publish-commands`. Also private: `_screenshots`, `_website-build`,
-# `_website-dev`, `_test-website` and `_uninstall-dotnet-10`.
+# `_website-dev`, `_test-website`, `_film` and `_uninstall-dotnet-10`.
 
 # ── OS detection ──────────────────────────────────────────────────
 # All recipes assume a POSIX shell. On Windows we use Git Bash (bundled with
@@ -145,7 +145,7 @@ KOVER_PERCENT = $(if $(COVERAGE_TOOLS_PREBUILT),$(DOTNET) target/coverage-tools/
 .PHONY: build ci test lint fmt clean audit setup _build-profiler-ui \
         reinstall-vsix install-vsix uninstall-vsix install-dotnet-10
 
-.PHONY: _screenshots _website-build _website-dev _uninstall-dotnet-10 \
+.PHONY: _screenshots _website-build _website-dev _film _uninstall-dotnet-10 \
         _package-vsix-linux-x64 _package-vsix-linux-arm64 \
         _package-vsix-darwin-arm64 _package-vsix-darwin-x64 \
         _package-vsix-win32-x64 _package-vsix-win32-arm64 \
@@ -805,6 +805,16 @@ _website-dev:
 	@echo "==> Starting website development server..."
 	npm run dev --prefix src/website
 
+# [WEB-DESIGN-MEDIA] Shoot and publish one website film from the packaged VSIX:
+# `make _film FILM=fsharp`, or `PUBLISH=1` to publish the last shoot again.
+# The films live in src/website/tools/profiler-film/films; its README says how.
+_film:
+	@test -n "$(FILM)" || { echo "usage: make _film FILM=<film> [PUBLISH=1]" >&2; exit 1; }
+	$(if $(PUBLISH),,$(MAKE) _build-vsix VSIX_PREBUILT=1)
+	@# Install once: npm ci would wipe the voice model kokoro-js caches in node_modules.
+	test -d src/website/tools/profiler-film/node_modules || npm ci --prefix src/website/tools/profiler-film
+	node src/website/tools/profiler-film/film.mjs $(FILM) $(if $(PUBLISH),--publish,)
+
 _test-website:
 	@echo "==> Running website Playwright tests..."
 	npm ci --prefix src/website

docs/plans/PROFILER-PLAN.md

@@ -186,6 +186,23 @@ These are SharpLsp's shipped profiler features. Cutover is blocked until each ro
   - The Rust `profiler_shared` tests raced: each nextest process rewrote the engine beside the shared server binary while another test's engine started, and lost the race with "exited before attaching". The engine is now installed atomically.
   - The suites' lifecycle and flows live once in `profiler-kit.ts` (`profilingSuite`, `profileFromRowFor`, `stopTracking`, `snapshotSession`, `retentionGraph`). The Deslop gate is ratcheted from 6.73 % to 6.65 %.
 - [x] The `profiler` chunk on macOS (`make _test-vsix-shard CHUNK=profiler`, 2026-10-15): **294 passing, 0 failing**, after the defects below were fixed; the debugger, manifest and configuration suites they touch pass too (30 passing).
+- [x] F# on real programs: `FsPricingEngine` (CPU: CSV parsing, discount rules, fuzzy search, routing, reports) and `FsOrderLedger` (an event handler, a memoize cache, a mailbox audit trail, undo history and rejection capture, each rooted at its own module value), under `src/sharplsp/tests/fixtures/fsharp-profiling/`. The suites `profiler-fsharp-cpu-e2e`, `-memory-e2e`, `-retention-e2e`, `-corner-e2e`, `-soak-e2e`, `-soak-cpu-e2e` and `profiler-lsp-protocol-e2e`, `-cpu-e2e`, `-memory-e2e` drive them through the UI and over the wire. The defects they found are fixed in lspkit and the host:
+  - idle samples counted as active, so a main thread waiting on stdin took half of every profile (lspkit#12);
+  - a frame every thread shares counted once per thread, so its samples outnumbered the ticks (lspkit#12);
+  - naming a runtime that cannot profile a busy PID answered `ALREADY_PROFILING`, not `UNSUPPORTED_RUNTIME` (lspkit#12);
+  - a `ResizeArray<Order>`'s `Order[]` painted the line that builds an `Order` (lspkit#12);
+  - a retention graph that left instances out still said it was complete (lspkit#12);
+  - a memory step on a program that had exited answered `OTHER`, not `PROCESS_NOT_FOUND` (lspkit#12);
+  - a second diff kept the first diff's leak badges (lspkit#12);
+  - FSharp.Core types were placed at its build server's paths, which do not exist here (lspkit#13);
+  - `objectsByType` dropped the requested type from its summary when it ranked below the twenty largest, as F# union cases do (lspkit#13);
+  - the retention graph never drew a root's label, so a module value's `static sessions` looked like any holder (lspkit#14);
+  - the timeline's peak fell after a release, because `peak_bytes` was each snapshot's own heap size (lspkit#15);
+  - a heap snapshot taken during a CPU session named types by bare address, because the CPU session enabled the runtime provider (lspkit#15);
+  - a hot function with no location reached the UI as file `""`, line 0 (lspkit#15);
+  - the UI's Stop never sent `memory.stop`, so the server kept the session and its graphs and still answered steps (lspkit#15);
+  - a timed CPU session never left `cpu.list` once its duration elapsed (lspkit#15);
+  - the host accepted a profiler command with no argument, two arguments or a non-object argument (`profiler/host.rs`).
 - [ ] Assertions of the deleted suites with no replacement: counters webview and counter sessions, `openTrace`/`convertTrace` of a `.nettrace`, dump collection with cancellable progress, `inspectObject`, `diffSnapshots`/`detectLeaks` panels over dumps, graph browser interactions. Each maps to a parity gap above.
 
 ### Documentation and cutover

docs/reviews/V0.22.0-STATIC-REVIEW.md

@@ -8,6 +8,18 @@ The .NET Framework and multi-targeting feature has been implemented across the R
 
 The spec explicitly excludes debugging the desktop .NET Framework CLR: Debug of a Framework-only target is refused, while Run uses the desktop CLR on Windows. “Finished” must be judged against that stated scope, not against availability of a .NET Framework debugger.
 
+## Resolution
+
+Re-checked against `main` on 2026-10-18 (`817d359f`):
+
+| Finding | Status |
+|---|---|
+| Parenthesized C# project name read as a framework | Fixed (#382). `FrameworkOf` accepts only the `(tfm)` suffix past the project file's stem. Covered by a real single-target `Utility(net48).csproj` and a multi-targeted `Tool(net10.0).csproj` in `WorkspaceManagerTargetFrameworkTests`. |
+| Switch announced before invalidation and diagnostics | Fixed (#383). The host clears `nav_cache`, publishes every open document's diagnostics and waits for each, then announces and asks for refreshes. `test_full_stack_target_framework_switch` asserts that order, using a warning only `net10.0` compiles. |
+| Run/F5 project paths compared with the wrong case rule | Already fixed before this re-check: `chosenFramework` uses `samePath` from `paths.ts`, and `run-debug-target.test.ts` covers the explicit-project, differently cased focused project on Linux for C# and F#. |
+| Explicit Windows-only selection passes with zero tests | Fixed (#384). Off Windows, a `MOCHA_FILES` pattern that names only `windowsOnly` suites fails with the platform reason (`test-runner-selection.test.ts`). |
+| Dependabot skips `detect-changes`, and the terminal gate passes the skipped run | Fixed. The skip is intended: bumps are tested once, at the `dependabot-upgrades → main` consolidation PR. But `CI` now FAILS a Dependabot PR to `main`, so a security bump the sweep did not retire cannot merge untested ([DIST-CI-DEPENDABOT], `security-gates.test.mjs`). |
+
 ## Findings, in priority order
 
 ### [P2] A single-target C# project name can be mistaken for a framework

docs/specs/DESIGN-SYSTEM.md

@@ -126,17 +126,18 @@ Page-level grids and marketing card styles must not leak into `.prose`. Do not r
 
 Screenshots and video of a SharpLsp feature show the real extension, never a mock-up.
 
-- **Capture.** The profiler media is shot by `src/website/tools/profiler-film/`. `capture.mjs` drives an isolated VS Code (its own user-data and extensions directories, never the user's editor) running the packaged VSIX against a real .NET app, and records stills and real-time screencasts. `render.mjs` turns them into the site's assets.
+- **Capture.** The profiler media is shot by `src/website/tools/profiler-film/` (its README says how). Every shot runs on `kit/set.mjs`'s `onSet`: an isolated VS Code (its own user-data and extensions directories, never the user's editor) running the packaged VSIX against a real .NET demo app from `demos/`. A walkthrough is one file in `films/`, a list of scenes that `film.mjs` shoots as real-time clips and publishes; `kit/acts/` tells the CPU and memory stories, so each film supplies only its words, its demo and its fix. `films/leak.mjs` films the C# Storefront demo and `films/fsharp.mjs` its F# twin, FsStorefront, so every F# claim the docs make about the profiler is shown on F# code. The composited one-minute film is `films/profiler/`: `capture.mjs` records stills and screencasts, and `render.mjs` turns them into the site's assets.
 - **Honesty.** Footage plays in real time unless the frame says otherwise: every fast-forwarded stretch carries an on-screen speed marker, and captions name time-lapses. Motion graphics (title cards, camera moves, highlights, the cursor) frame the footage but never alter what the editor shows.
 - **Privacy.** No frame may show a user name, home directory, hostname or terminal. The demo app runs from a neutral temporary path and the terminal is never opened.
 - **Formats.**
   - Films ship as H.264 MP4 and VP9 WebM with a WebP poster.
   - Stills ship as WebP, 2400px wide, with `width` and `height` set.
-  - Marketing pages may autoplay a film only when it is muted and looping; docs pages use controls.
+  - Marketing pages may autoplay a film only when it is muted and looping; docs pages use controls and are never muted.
   - Every autoplaying video pauses and shows its controls under `prefers-reduced-motion: reduce`.
   - Every video ships a WebP poster, the films and the recordings alike.
+- **Narration.** A walkthrough may be narrated (`kit/sound/`). Its reviewed script (`films/<film>.narration.json`) is spoken by Kokoro-82M through `kokoro-js` on the capturing machine, never by a speech service or the operating system's voices, whose licences forbid publishing. Each line starts just after its scene's cut, and a scene holds its last frame until its line ends. The music is an original score synthesised from code (`kit/sound/score.mjs`), so no film carries licensed or sampled music, and it ducks under the voice. A narrated film ships AAC in its MP4 and Opus in its WebM, plays unmuted with controls, and carries WebVTT captions in every site language (`<id>.<lang>.vtt`, the page's own language shown by default), listed under `captions` in the catalogue.
 - **One catalogue.** Every video on the site is declared once in `src/website/src/_data/videos.json`: encodings, size, length, upload date, and its label and caption in each language. Pages place a video with the `{% video "<id>" %}` shortcode (`src/website/video-figure.js`) and never repeat `<video>` markup. The shortcode renders in the page's language and emits a schema.org `VideoObject` whose name is the player's label, so search results describe exactly what a reader hears.
-- **Placement.** A video goes wherever a page talks about what it shows: the home page (the profiler film autoplaying in the hero, the CPU recording in the profiler section), the Getting Started page (the tour), the profiler docs (the film and both recordings), and every blog post that discusses the profiler or the editor experience, in every language the page is published in.
+- **Placement.** A video goes wherever a page talks about what it shows: the home page (the profiler film autoplaying in the hero, the CPU recording in the profiler section), the Getting Started page (the tour), the profiler docs (the film and both recordings), the F# profiling docs and the F# language page (the F# walkthrough), and every blog post that discusses the profiler or the editor experience, in every language the page is published in.
 
 ## [WEB-DESIGN-RESPONSIVE] Responsive Behavior
 

docs/specs/DISTRIBUTION-SPEC.md

@@ -867,10 +867,10 @@ inheriting a skip:
 
 1. Any upstream job that FAILED or was CANCELLED fails `CI`, and with it the merge.
 2. A SKIPPED upstream job satisfies it only when no code changed. That is how a
-   docs-only PR (every job skipped by [DIST-CI-CLASSIFICATION]) and a Dependabot
-   PR (the whole pipeline skipped, [DIST-CI-DEPENDABOT]) pass. When code changed,
-   `build`, every test leg and `coverage` MUST have succeeded; a skip among them
-   fails `CI` (#365).
+   docs-only PR (every job skipped by [DIST-CI-CLASSIFICATION]) passes. When code
+   changed, `build`, every test leg and `coverage` MUST have succeeded; a skip
+   among them fails `CI` (#365). A Dependabot PR, whose whole pipeline skipped
+   ([DIST-CI-DEPENDABOT]), always fails `CI`: it ran nothing.
 3. Every PHASE 4 leg and `coverage` states its condition explicitly:
    `!cancelled()`, code changed, and success of the jobs it consumes. GitHub's
    implicit `success()` spans the whole dependency chain, so a legitimately
@@ -929,7 +929,9 @@ superseded, and MUST NOT merge anything into `main` unattended.
    staged batch is verified once, by the full pipeline and CodeQL, on the
    `dependabot-upgrades → main` consolidation PR — the only route a dependency
    change takes into `main`. A staged security fix reaches `main` when that PR
-   is opened, so it is opened promptly.
+   is opened, so it is opened promptly. `CI` enforces the route: it FAILS a
+   Dependabot PR to `main`, so a security bump whose sweep did not retire it
+   cannot merge untested.
 4. `dependabot-automerge.yml` MUST exist on both branches: for `pull_request`,
    GitHub runs the workflow from the PR's BASE branch.
 
@@ -1107,6 +1109,10 @@ Invariants:
   distinction the manifest carries is `"linuxOnly": true` or
   `"windowsOnly": true`, which drops a chunk from the other platform's matrix. A
   chunk MUST NOT be declared in CI YAML.
+- **A selection never runs nothing.** The in-host runner fails a run whose
+  `MOCHA_FILES` pattern matches no compiled suite, and, off Windows, one whose
+  pattern names only suites of `windowsOnly` chunks: dropping every one would
+  report a green run of zero tests (`test-runner-selection.test.ts`).
 - **One runner, one shard target.** `make _run-vsix-suite` is the single recipe;
   `CHUNK` selects the slice (empty runs every suite) and `VSIX_SUITE_PREBUILT`
   says the suite is already compiled. Coverage is NOT a knob — the runner always

docs/specs/DOTNET-FRAMEWORK-SPEC.md

@@ -64,13 +64,16 @@ built at load or at the switch that needs them
 - `sharplsp/targetFramework` `{ textDocument }` → `{ active, available, project }`;
   `sharplsp/setTargetFramework` `{ textDocument, targetFramework }` switches the whole PROJECT in
   both languages and answers the same shape. A single-target project answers
-  `{ active: null, available: [] }`; a document no loaded project compiles answers
-  `{ active: null, available: [], project: null }`, and switching it is an error.
+  `{ active: null, available: [] }`, even one whose name ends in `(tfm)`: a framework is
+  only the suffix `MSBuildWorkspace` appends to the project FILE's stem. A document no
+  loaded project compiles answers `{ active: null, available: [], project: null }`, and
+  switching it is an error.
 - After a switch the host clears its navigation cache, re-publishes every open document's
-  diagnostics, sends `sharplsp/targetFrameworkChanged` `{ textDocument, active, available,
-  project }`, then each of `workspace/semanticTokens/refresh`, `workspace/inlayHint/refresh`,
-  `workspace/codeLens/refresh` and `workspace/diagnostic/refresh` whose `refreshSupport` the
-  client declared.
+  diagnostics and waits for each first publication (a failed or provisional fetch goes on
+  retrying in the background), sends `sharplsp/targetFrameworkChanged` `{ textDocument,
+  active, available, project }`, then each of `workspace/semanticTokens/refresh`,
+  `workspace/inlayHint/refresh`, `workspace/codeLens/refresh` and
+  `workspace/diagnostic/refresh` whose `refreshSupport` the client declared.
 - VS Code: a status-bar item `$(versions) <tfm>` shows the focused document's active framework
   when its project is multi-targeted, and runs `sharplsp.selectTargetFramework [tfm]`. Without
   an argument the command opens a pick placeheld `Target framework for <Project>` (file name, no

docs/specs/PROFILER-SPEC.md

@@ -62,7 +62,8 @@ flowchart LR
   - the record, the union case and the generic record each anchor, and are painted, at the exact `Program.fs` line that constructs them;
   - the dashboard names the union case `HotLeakAppFs.PayloadKind+Small`;
   - the diff scores the record as a leak;
-  - the leak's retention root is the module value `static items`, with no compiler-mangled name.
+  - the leak's retention root is the module value `static items`, with no compiler-mangled name;
+  - the fixture's line protocol builds the rest of the F# heap zoo one shape at a time: orders over a single-case union, a union case with fields, a list and a struct record; quotes behind a never-evicting cache; event-handler closures; a reference cycle; and arrays. Each anchors at its constructing line except arrays, which have none, and the growing record reaches Low confidence.
 - File-system paths go only through `src/editors/vscode/src/paths.ts`.
 
 ### [PROFILER-SHARED-LIBRARY-ENGINE] Engine bundling
@@ -651,7 +652,7 @@ Commands used per node:
 
 ### [PROFILER-GRAPH-WEBVIEW] Interactive Graph Webview
 
-The object graph renders as an interactive force-directed graph in a VS Code webview panel.
+The object graph renders as a layered graph in a VS Code webview panel: GC roots on top, named by their labels, and the target objects along the bottom ([PROF-UI-REFGRAPH] in lspkit).
 
 #### [PROFILER-GRAPH-WEBVIEW-LAYOUT] Graph Layout
 
@@ -762,7 +763,7 @@ output_directory = ".sharplsp/profiles"
 | Command palette | Start/stop trace, start/stop counters, collect dump, leak detection |
 | Webview panel | Display counter values as live-updating table |
 | Webview panel | Display heap stats as sortable table |
-| Webview panel | Interactive object retention graph (D3.js force-directed) |
+| Webview panel | Object retention graph, layered from GC roots down to the targets |
 | Webview panel | Heap snapshot diff table with growth indicators |
 | Quick pick | Process selection from discovered .NET processes |
 | File open | Open `.speedscope.json` output in browser/SpeedScope viewer |