v1.40.1.0 fix: refresh gbrain installer pin

[hnshah]

Summary

• Refresh bin/gstack-gbrain-install from the original gbrain v0.18.2 pin to current verified gbrain v0.35.4.0 (0c6fcab).
• Fix the immediate clean-first-run sync crash with current gbrain by accepting the gbrain sources list --json object shape ({ sources: [...] }) during hostname-fold migration planning.
• Normalize gstack-gbrain-detect version output from gbrain 0.35.4.0 to 0.35.4.0.
• Teach /setup-gbrain local PGLite paths to use ollama:nomic-embed-text automatically when no OPENAI_API_KEY is present and the local Ollama embedding probe succeeds.
• Regenerate/update /setup-gbrain docs and move the gstack-memory-ingest: CLI call format incompatible with every released gbrain (+ note on /gstack-upgrade pin staleness) #1305 part 2 TODO to completed for v1.40.1.0.

Why

Dogfooding the gstack/gbrain setup path surfaced that fresh /setup-gbrain installs were still pinned to the original v0.18.2-era gbrain even though current gstack memory features expect newer gbrain behavior. After refreshing the pin, clean /sync-gbrain --code-only hit a compatibility crash before registering the repo source:

gstack-gbrain-sync fatal: list.find is not a function

That came from one helper still assuming the old array-shaped sources list while gbrain v0.35.4.0 returns { sources: [...] }.

The same dogfood run also showed first-sync embedding failures on a machine with no OPENAI_API_KEY, even though gbrain already supports local Ollama embeddings. The setup skill now probes the actual local embedding path and initializes PGLite with --model ollama when it is already working.

Test plan

• bun run gen:skill-docs --host all
• bun test test/gbrain-detect-install.test.ts (15 pass)
• bun test test/gstack-gbrain-sync.test.ts test/gbrain-detect-shape.test.ts test/gbrain-detect-install.test.ts (59 pass)
• bun test test/setup-gbrain-path4-structure.test.ts test/gen-skill-docs.test.ts
• HOME=$(mktemp -d) gbrain init --pglite --json --model ollama verified config writes embedding_model: "ollama:nomic-embed-text", embedding_dimensions: 768
• git diff --check

Dogfood

Clean throwaway HOME + GSTACK_HOME:

• gbrain init --pglite --json
• gstack-gbrain-detect -> gbrain_version: "0.35.4.0", gbrain_local_status: "ok"
• gstack-gbrain-sync --dry-run --code-only
• gstack-gbrain-sync --code-only
• Added a tiny hello.js, reran sync, and source page_count reached 1.
• Confirmed local Ollama embedding probe succeeds: gbrain providers test --model ollama:nomic-embed-text -> 768 dims.

Verification

• Verified 0c6fcab555b1ca0de80f47dbc2bd692499a82590 is current garrytan/gbrain master.
• Verified that commit's package.json reports 0.35.4.0.

[hnshah]

CI note: the code/docs gates are passing (check-freshness, version gate, actionlint, macOS/Ubicloud gates, Windows tests). The remaining build-image failure is from the workflow trying to push the CI image to GHCR from a fork: denied: installation not allowed to Write organization package.

[hnshah]

Dogfood update: after refreshing the installer pin to current gbrain, I ran a clean HOME/GSTACK_HOME first-run flow and found the next compatibility bug immediately: gstack-gbrain-sync --code-only crashed with list.find is not a function because sourceLocalPath() still assumed an array-shaped gbrain sources list --json, while current gbrain returns { sources: [...] }.

Amended this PR to:

• accept both the old array shape and current object shape in sourceLocalPath()
• add a regression test for the current object shape
• normalize gstack-gbrain-detect version output from gbrain 0.35.4.0 to 0.35.4.0

Verification:

• bun test test/gstack-gbrain-sync.test.ts test/gbrain-detect-shape.test.ts test/gbrain-detect-install.test.ts — 59 pass
• clean dogfood sync now registers the repo source; after adding a tiny hello.js, page_count reached 1

[hnshah]

Follow-up from dogfooding local embeddings: this now keeps fresh local PGLite setup on the local path when possible. If OPENAI_API_KEY is absent and gbrain providers test --model ollama:nomic-embed-text succeeds, the setup skill runs gbrain init --pglite --json --model ollama.

I verified the command writes embedding_model: "ollama:nomic-embed-text" and embedding_dimensions: 768 in a throwaway HOME.

[hnshah]

Updated this PR to the latest ready gbrain release in de176a5a.

What changed:

• bin/gstack-gbrain-install now pins garrytan/gbrain at 03947665e4dbfeaf8a5542d160a0f4b89e4ae747 / v0.36.0.0.
• Updated the dry-run pin assertion.
• Updated /setup-gbrain generated/template example copy.
• Updated the completed TODO note to reflect v0.36.0.0.

Verification:

• Confirmed upstream gbrain master package version is 0.36.0.0 at 03947665e4dbfeaf8a5542d160a0f4b89e4ae747.
• bun test test/gbrain-detect-install.test.ts test/setup-gbrain-path4-structure.test.ts test/gen-skill-docs.test.ts — 420 pass.
• bun run gen:skill-docs --host all
• git diff --check
• Clean no-existing-clone dry run prints the new pinned commit/tag.

[hnshah]

CI update after de176a5a:

Passing:

• check-freshness
• version gate
• actionlint
• macOS gate
• Ubicloud gate
• Windows tests
• E2E report

Only failing:

• build-image, same fork/package-permission failure as before: denied: installation not allowed to Write organization package while pushing ghcr.io/garrytan/gstack/ci.

The Docker image builds successfully and fails only at the GHCR push step.