Memory: pasclaw memory provision + onboarding download hook

[FMXExpress]

Summary

PR #165's PasClaw.Memory.Vector.Open() needs three artifacts under $PASCLAW_HOME/cache/localvector/ — sqlite-vec extension, ONNX Runtime, MiniLM model + vocab — before the hybrid backend will engage. Until now operators had to know to download those themselves. This PR closes that loop with one new command and an onboarding hook.

New command: src/cmd/PasClaw.Cmd.Memory.pas

Two subcommands:

pasclaw memory provision

Downloads sqlite-vec, ONNX Runtime, MiniLM model + vocab into $PASCLAW_HOME/cache/localvector/. Idempotent — re-running skips anything already present and redownloads anything suspiciously small (treated as a partial fetch by TModelDownloader, mirrors the localvector convention).

Calls the in-tree provisioning helpers that PR #165 brought in:

Helper	What it fetches
LocalVector.VecProvision.EnsureVec0	sqlite-vec from asg017/sqlite-vec GitHub releases — cross-platform: linux / macos / windows × x86_64 / aarch64
LocalVector.OrtProvision.EnsureOnnxRuntime	ONNX Runtime; auto-download is win-x64 only per CanAutoProvisionRuntime. Other platforms print install instructions for the system package manager.
LocalVector.Downloader.TModelDownloader	MiniLM model.onnx (~90 MB) + vocab.txt from HuggingFace

Per-step status: green ✓ on success, red ✗ on failure, dim · for "platform doesn't support auto-download here". Final summary tells the operator whether the hybrid backend is ready or whether memory_search will still fall back to FTS. Exit code 0 on full success, 1 if any of the three steps failed.

pasclaw memory status

Read-only view: shows the active backend (hybrid when vector_search_enabled and artifacts present, else FTS-only), the cache directory, and a per-artifact ✓/· checklist mirroring what PasClaw.Memory.Vector.Open looks for. Useful for "did the provision actually take" diagnostics without running a real memory_search.

Onboarding hook

After the existing PromptVectorSearch question, when the operator said yes, ask once more: "Download now? [y/N]". Default NO because the artifacts are ~91 MB and an Enter-through-everything onboarding shouldn't trigger a fat download the user didn't opt into. Yes invokes Cmd_Memory_Run(['provision']) inline; No prints "deferred, run pasclaw memory provision when ready".

Dropped the now-stale "follow-up release downloads the model on first use" disclaimer from the prompt — that follow-up release is this PR.

Test plan

Verified end-to-end on Linux/FPC:

$ pasclaw memory status
  · sqlite-vec missing at .../vec0.so
  · all-MiniLM-L6-v2 model missing at .../model.onnx
  · vocab missing at .../vocab.txt

$ pasclaw memory provision
  ✓ sqlite-vec extension installed at .../vec0.so       (160 KB)
  · ONNX Runtime auto-download is win-x64 only on this build;
    install via your system package manager:
      Debian / Ubuntu:  apt install libonnxruntime-dev
      macOS (Homebrew): brew install onnxruntime
      Other Linux:      see https://onnxruntime.ai/docs/install/
  ✗ ONNX Runtime: Could not load onnxruntime. Put onnxruntime.so
    next to the executable or on the loader path.
    (install per the instructions above, then re-run
     `pasclaw memory provision`)
  ✓ embedding model all-MiniLM-L6-v2 (~90 MB) installed at .../models/...
  ! provisioning incomplete — memory_search will fall back to FTS-only

$ pasclaw memory status
  ✓ sqlite-vec at .../vec0.so
  ✓ all-MiniLM-L6-v2 model at .../model.onnx
  ✓ vocab at .../vocab.txt

$ NewVectorMemoryIndex.Open()  → returns False (ONNX missing → falls back
                                  to FTS-only as designed)

• Real download of sqlite-vec (160 KB tar.gz extracted to vec0.so).
• Real download of MiniLM model.onnx (90 MB) + vocab.txt (232 KB) from HuggingFace.
• ONNX-Runtime-missing path prints install instructions cleanly and exits 1.
• pasclaw memory status reports the post-provision state accurately.
• PasClaw.Memory.Vector.Open() still returns False quietly when ONNX is missing — fallback to FTS works.
• Full test suite green: smoke + hashline + toolview + anthropic-server-tools + openai-server-tools + println-helper + utf8-codepage-tag + gemini-schema-strip + markdown-render + json-utf8-roundtrip.
• Live pasclaw memory provision on Windows-x64 to confirm the ONNX auto-download branch (didn't fire on the Linux test rig).
• Live exercise of the hybrid memory_search path on a host with libonnxruntime installed system-wide — that's the full integration test which needs an environment we don't have in CI.

Stacks on top of

PR #165 (the in-tree localvector port). #165 must merge first.

---

Generated by Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 71f37eeff0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Report ONNX availability in memory status

When pasclaw memory status runs on a host that has vec0 and the model/vocab files but no loadable ONNX Runtime (common on Linux/macOS, where this command does not auto-download it), this checklist prints every listed artifact as present even though TVectorMemoryIndex.Open still calls EnsureOnnxRuntime(..., False, ...) before enabling the hybrid backend. That makes the diagnostic say the provision “took” while memory_search will continue falling back to FTS-only; include an ONNX/cache-or-loader check here so status mirrors the actual backend gate.

Useful? React with 👍 / 👎.