fix(schema-pack): embed bundled pack YAMLs in the compiled binary (+ entity-slug orphan floor)

[JiraiyaETH]

The bug (Layer 0 — the foundational outage)

Bundled schema-pack YAMLs are never embedded in bun build --compile binaries. The pack locators resolve paths via fileURLToPath(import.meta.url) + existsSync, which points at /$bunfs/root/... in a compiled binary where the YAML is not on disk. Symptoms in a deployed binary:

• gbrain schema active → unknown schema pack: gbrain-base. Any active pack that extends: gbrain-base fails to resolve because the bundled parent can't load, taking the whole resolution down.
• put_page silently degraded — the pack-load try/catch swallowed the failure, reverted to a hardcoded prefix table, and skipped type validation → accumulating page-type sprawl.
• gbrain schema show gbrain-base and the lens packs (gbrain-creator/investor/engineer/everything) were unreachable from the CLI/MCP.

build:schema is not the cause (it regenerates Postgres DDL); a plain rebuild reproduces the bug. The fix is a source-level embed.

The fix

• New central registry src/core/schema-pack/bundled-packs.ts imports all 7 bundled pack YAMLs via import … with { type: 'file' } (Bun embeds the bytes; the import evaluates to a readable path in both bun run and --compile). Mirrors the existing WASM-embed pattern (src/core/chunkers/code.ts + scripts/check-wasm-embedded.sh). Exports BUNDLED_PACK_PATHS, BUNDLED_PACK_LIST, BUNDLED_PACK_NAMES, bundledPackPath().
• Routed every pack-load surface through it (the previous code had four divergent locators): load-active.ts defaultPackLocator, the CLI schema show/validate/list/use/fork (packPathByName + runList), MCP list_schema_packs + schema_lint named-pack, and the read-only mutability set in mutate.ts (was 3 names, now all 7).
• Fail-closed put_page — a configured pack (resolution source ≠ default) that won't load now throws instead of silently degrading; a default-pack load failure loud-warns.
• Functional regression guard scripts/check-packs-embedded.sh + scripts/packs-smoketest.ts: compiles a probe and loads every bundled pack through the compiled binary (not strings | grep). Wired into check:all, run-verify-parallel, and ci-local. Updated the BUNDLED_PACK_NAMES size test 3 → 7. Docs refreshed.

Layer 1a — entity-slug floor (no literal-null/hyphen-flattened fact orphans)

• extract.ts: a model-emitted entity of "null"/"none"/whitespace is coerced to JSON null (the fact stays unbound, never a null-slug orphan).
• entities/resolve.ts: non-entity tokens → null binding (kept unbound via the existing legacy bucket, not dropped); the floor uses slugifyEntityPath, which preserves an explicit entity-prefix path (companies/Acme Co → companies/acme-co) but flattens non-prefix slashes (A/B Partners → a-b-partners) so it can't mint arbitrary nested pages past the stub guard.
• Read path (recall + MCP list_facts): a null resolution returns no facts instead of querying the raw string and surfacing legacy orphan rows.

Deferred (called out, not in this PR)

• registry.ts resolvePack returns the child manifest unchanged for extends/borrow_from, so gbrain-recommended/gbrain-everything resolve with 0 page types. Not the outage (operational packs declare their types directly), but worth a follow-up to honor the lens-pack design.
• Deriving the entity prefix set from the active pack's path_prefixes instead of the hardcoded people/+companies/.

Tests

typecheck clean · scripts/check-packs-embedded.sh passes (compiled-binary probe) · the schema-pack/facts/entity suites pass (246 tests). Verified live: a freshly compiled binary resolves the configured active pack and schema validate <lens-pack> reads its embedded /$bunfs/root/...yaml.

🤖 Generated with Claude Code