feat: auto-fallback to SQLite when embedded SeekDB is unavailable

[knqiufan]

Summary

• Add powermem.platform_defaults with choose_default_database_provider() to centralize zero-config storage selection.
• When DATABASE_PROVIDER and OCEANBASE_HOST are unset:
  • Non-Linux platforms default to SQLite (./data/powermem_dev.db).
  • Linux probes embedded SeekDB (pyobvector, pyseekdb, pylibseekdb); uses OceanBase embedded mode when available, otherwise falls back to SQLite with a WARNING.
• Wire the helper into config loading, Memory/CLI setup, OceanBase embedded init, and server startup.
• Expose storage_capabilities on /api/v1/system/status so clients can see SQLite limitations (no Graph Store, sub_stores, sparse vectors, SkillStore).
• Update .env.example / .env.example.full comments to document platform-aware defaults.

Closes #992

Test plan

• pytest tests/unit/test_storage_default_provider.py
• pytest tests/unit/test_seekdb_default_storage.py
• Windows zero-config: powermem-server starts, memory APIs return 200 (not 503)
• Linux + powermem[seekdb]: zero-config still uses embedded SeekDB at ./seekdb_data
• Explicit DATABASE_PROVIDER=oceanbase on non-Linux still surfaces a clear validation error (no silent override)

[wayyoungboy]

#1031

[wayyoungboy]

Thanks for the update. I rechecked this after the latest commits. The checks are green, but GitHub currently reports this branch as conflicting with main, so I cannot mark it LGTM yet.

Please rebase or merge the latest main and ping me; I will review the rebased diff again.

[knqiufan]

Thanks @wayyoungboy. Rebased on the latest main (ffefdb2).

The only conflict was a non-semantic import ordering in src/powermem/config_loader.py — kept both the new NoopConfig import (from main) and the platform_defaults imports (from this PR). Verified noop is still registered in BaseLLMConfig._registry after the rebase.

Test plan still green:

• pytest tests/unit/test_storage_default_provider.py — 7/7 passed
• pytest tests/unit/test_seekdb_default_storage.py — 3/3 passed

Ready for re-review whenever you have a moment.

[wayyoungboy]

Thanks for rebasing. I rechecked the latest head (50a9b5c). The previous merge conflict is resolved and the GitHub checks are green, but I still see a few items to fix before LGTM:

1. The new default-provider test is not hermetic, and it exposes an explicit-vs-default mismatch for env-file based config. DatabaseSettings reads the auto-detected settings env file through settings_config() / _DEFAULT_ENV_FILE, but default_database_provider() and database_provider_explicitly_configured() only inspect os.environ. With a temporary HOME containing .powermem/.env and DATABASE_PROVIDER=oceanbase, DatabaseSettings().provider resolves to oceanbase while default_database_provider() resolves to sqlite, so tests/unit/test_seekdb_default_storage.py::test_database_settings_default_provider_matches_platform_helper fails. Please either make the test isolate settings env-file loading, or make the platform default/defaulted helpers use the same source of truth as DatabaseSettings.

2. /api/v1/system/health is public, but it now returns raw startup_error text from service initialization. Those exceptions can include deployment details such as dependency errors, local paths, or database connection information. Please keep the public health response coarse/sanitized, and expose detailed startup errors only via authenticated status endpoints or logs.

3. git diff --check origin/main...HEAD currently fails with src/server/utils/service_errors.py:19: new blank line at EOF.

Validation I ran:

• Activated .venv and ran python -m pytest tests/unit/test_storage_default_provider.py tests/unit/test_seekdb_default_storage.py -q: 1 failed, 9 passed.
• Activated .venv and ran the same tests with a temporary clean HOME: 10 passed.
• Activated .venv and ran python -m pytest tests/unit/test_config_loader.py tests/unit/test_cli_config_show.py -q with a temporary clean HOME: 9 passed.
• Activated .venv and ran python -m pytest tests/unit/server -q with a temporary clean HOME: 44 passed.
• Activated .venv and ran python -m pip check: no broken requirements.

[knqiufan]

Thanks @wayyoungboy. I addressed the three review items in the latest push (2a70a0f).

1. The default-provider helpers now use the same configuration sources as DatabaseSettings: process env first, then the configured env files (POWERMEM_ENV_FILE / _DEFAULT_ENV_FILE). I also updated the defaulted calculation to use the same helper, so env-file based DATABASE_PROVIDER / OCEANBASE_HOST no longer diverges from DatabaseSettings.

2. /api/v1/system/health no longer exposes raw startup exceptions. It only returns coarse readiness information; detailed startup errors remain available from the authenticated /api/v1/system/status endpoint and server logs.

3. Fixed the trailing blank line reported by git diff --check.

Validation:

• python -m pytest tests/unit/test_storage_default_provider.py tests/unit/test_seekdb_default_storage.py -q
• python -m pytest tests/unit/server/test_system_health.py -q
• python -m pytest tests/unit/test_config_loader.py tests/unit/test_cli_config_show.py -q
• git diff --check origin/main

Please re-review when you have a chance.

[wayyoungboy]

Rechecked the current PR state after the latest author update. GitHub currently reports this branch as conflicting with main, so final review/LGTM is still blocked until the branch is rebased or merged with the latest base branch.

The existing requested-changes review remains the active review state for now. After the conflict is resolved, please ping for a fresh review of the rebased diff and updated test results.

[knqiufan]

@wayyoungboy Thank you for the careful recheck and review.

I merged the latest main into this branch and pushed the conflict resolution in 6d4375c.

What changed in this update:

• Resolved the apps/claude-code-plugin/SETUP.md conflict by keeping the latest server/MCP extras wording while preserving this PR's SQLite fallback note for platforms where embedded SeekDB is unavailable.
• Rechecked the previous review items after the merge: env-file based default-provider resolution still uses the same source of truth as DatabaseSettings, and public /api/v1/system/health still returns only coarse readiness without exposing raw startup errors.
• Fixed a Windows-specific server bind probe issue found while rerunning the server tests: the probe no longer sets SO_REUSEADDR on Windows, where it can incorrectly allow binding to an actively listened port.
• Cleaned the trailing whitespace caught by git diff --check in the rebased/merged diff.

Validation:

• python -m pytest tests/unit/test_storage_default_provider.py tests/unit/test_seekdb_default_storage.py -q -> 14 passed
• python -m pytest tests/unit/server/test_system_health.py -q -> 1 passed
• python -m pytest tests/unit/test_config_loader.py tests/unit/test_cli_config_show.py -q -> 9 passed
• python -m pytest tests/unit/server -q -> 74 passed, 1 skipped
• python -m pytest tests/unit/intelligence/test_importance_evaluator.py -q -> 17 passed
• python -m pip check -> no broken requirements
• git diff --check origin/main...HEAD -> clean

Please take another fresh look when convenient.

[wayyoungboy]

Rechecked the current head after the latest merge/rebase. The previously requested items appear addressed in the diff:

• The platform default helper now reads the same explicit provider sources as DatabaseSettings (process env first, then configured env files), so env-file based DATABASE_PROVIDER / OCEANBASE_HOST no longer diverges from the default-provider decision.
• Public /api/v1/system/health no longer exposes raw startup errors; the added test checks that sensitive-looking startup details are absent from the public health response.
• The native dependency probe now handles modules with broken / missing specs as unavailable, which covers the CI stability issue around storage probes.
• The latest status checks are green and GitHub reports the branch as mergeable.

I did not find a new blocking issue in this static re-review. I am not turning this into an approval in this pass because I did not run the PR test plan locally.

[lightzt99]

LGTM