fix(lakebase): follow next_page_token when listing Postgres projects

[karen-amaya]

Summary

_lookup_via_postgres_api calls GET /api/2.0/postgres/projects once without following next_page_token. That endpoint paginates at 10 projects/page, so in workspaces with more than 10 Postgres projects, an project past page 1 is invisible to this lookup — causing the misleading error "No Lakebase Autoscaling endpoint matched... legacy Provisioned instances are not supported", even with a valid, correctly-permissioned project.

Confirmed in production (2026-09-04): the project was on page 2 of 19. Fix: loop over next_page_token until exhausted.

Linked Issue / milestone

Closes #160

Plan

N/A — single-file fix, no formal plan doc for this one.

Type of change

• feat — new feature
• fix — bug fix
• docs — documentation only
• refactor — no behaviour change
• test — tests only
• perf — perf improvement
• ci / build / chore — tooling

Author checklist

• Conventional Commit PR title (<type>(<scope>): <subject>).
• changelogs/<today>.log updated with title + context + numbered changes + files + test result.
• Tests added or modified for every behaviour change.
• uv run pytest tests/<scope>/ green locally.
• pre-commit run --all-files clean (or skipped hooks documented).
• If src/agents/** or any MLflow-traced LLM path changed:
  • SPEC.md present in .planning/<slug>/.
  • tests/eval/datasets/<agent>/dataset.jsonl has ≥20 examples (new) or ≥10 (change).
  • MLflow eval run URI in PR body below.
  • Judge score ≥ baseline + delta, or explicit waiver here.
• If src/mcp-server/** changed: uv run pytest tests/mcp/ -m mcp green.
• No gsd-* references re-introduced.

MLflow eval run

Test plan

• uv run pytest tests/ -m "not e2e and not property and not eval" --cov-fail-under=90
• Per-package coverage thresholds (scripts/check_coverage.py) green for touched packages.
• If new MCP tool: uv run pytest tests/mcp/integration/test_tool_schemas.py includes it.

Reviewer hint

See docs/PR_REVIEW_CHECKLIST.md. Numbered items map 1:1 to comments — #3: missing OntoBricksError subclass for new condition is more useful than "fix error handling".

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[benoitcayladbx]

Thanks for the diagnosis and the patch — the next_page_token loop in _lookup_via_postgres_api is the right idea, and the production note (project on page 2 of 19) is exactly the failure mode.

This PR is superseded. The same bug is already fixed on the maintained branches (not on master, which is why GitHub still shows #160 open):

• develop: fa617ca
• 0.9.0: 0fb7456

Please close this PR rather than merging it onto master. The landed fix is a bit broader than this diff:

1. Shared helper — LakebaseProjectService.list_projects follows every next_page_token and sets page_size=100 (API max). This PR keeps the default ~10/page, so a 19-project workspace still needs two round-trips; fine, but we wanted one shared loop.
2. Same one-shot GET in two other callers — Settings (graph_engine_lakebase_projects_result) and provisioning (LakebaseProvisioner._resolve_project_path) had the identical truncation. Auth-only pagination would still leave the dropdown and provisioner broken in large workspaces.
3. Misleading error — the “legacy Provisioned instances are not supported” copy is unchanged here. After a full walk with no host match, the landed code reports that no listed Autoscaling endpoint matched PGHOST.
4. Tests / changelog — no regression for a page-2 host, and the author checklist changelog/tests boxes are still unchecked.

Happy to take a follow-up if anything in the landed helper doesn’t cover your workspace. Closing #160 against the develop / 0.9.0 commits.