fix(proxy): support Pro Lite account routing

[aruis]

Summary

• Recognize upstream prolite account plan values without rewriting stored account rows.
• Map Pro Lite accounts to Pro-equivalent primary and secondary usage capacities for dashboard totals and capacity-weighted routing.
• Treat Pro Lite as Pro-equivalent when model registry plan gates allow pro, so Pro-gated models can select Pro Lite accounts.
• Add OpenSpec change artifacts and regression coverage for auth normalization, usage summaries, dashboard overview, and proxy account selection.

Fixes #557

Validation

• uv run pytest -q tests/unit/test_auth.py tests/unit/test_usage.py tests/unit/test_proxy_load_balancer_refresh.py tests/integration/test_dashboard_overview.py
• uv run ruff check app/core/plan_types.py app/core/usage/__init__.py app/modules/proxy/load_balancer.py tests/unit/test_auth.py tests/unit/test_usage.py tests/unit/test_proxy_load_balancer_refresh.py tests/integration/test_dashboard_overview.py
• uv run ty check app/core/plan_types.py app/core/usage/__init__.py app/modules/proxy/load_balancer.py tests/unit/test_auth.py tests/unit/test_usage.py tests/unit/test_proxy_load_balancer_refresh.py tests/integration/test_dashboard_overview.py

Notes

• uv run openspec validate --specs could not run locally because the openspec executable is not installed in this environment.
• A local amd64 Docker test image was exercised against the reported Pro Lite account, confirming dashboard quota totals and gpt-5.4 request routing after proxy configuration was corrected.

[Soju06]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5419de7772

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]

Preserve direct matching for unknown account plans

account_plan_matches_allowed now rejects any plan string that is not in ACCOUNT_PLAN_TYPES because normalize_account_plan_type returns None, so accounts with a new upstream plan (for example, any future plan value) will never pass model gating even when the model registry explicitly allows that same plan. This is a regression from the previous _filter_accounts_for_model behavior (a.plan_type in allowed_plans) and can cause modeled requests to fail with NO_PLAN_SUPPORT_FOR_MODEL until code is updated for each new plan string; keep a fallback direct membership check before applying equivalence mapping.

Useful? React with 👍 / 👎.

[Soju06]

Thanks @aruis -- the prolite plumbing looks right and Fixes #557 is on the money. Codex review flagged one real regression that I think we should address before merge:

P1 -- account_plan_matches_allowed rejects unknown plans

The old _filter_accounts_for_model did a direct string comparison:

allowed_plans = ...  # from upstream model.available_in_plans
return [a for a in accounts if a.plan_type in allowed_plans]

The new helper short-circuits to False when normalize_account_plan_type(value) returns None:

def account_plan_matches_allowed(value, allowed_plans):
    normalized = normalize_account_plan_type(value)
    if normalized is None:
        return False  # <-- regression
    ...

normalize_account_plan_type only knows the values currently in ACCOUNT_PLAN_TYPES. So if OpenAI later ships a new plan slug (the next prolite-style addition), accounts on that plan will silently fail model gating for every model, including ones whose registry entry explicitly lists the new plan in available_in_plans. This is the regression direction.

Suggested shape -- preserve a direct string fallback so unknown-but-explicitly-allowed plans still work:

def account_plan_matches_allowed(value, allowed_plans):
    cleaned = _clean_plan_type(value)
    if not cleaned:
        return False
    normalized = canonicalize_account_plan_type(cleaned)
    normalized_allowed = {plan.lower() for plan in allowed_plans}
    candidate = (normalized or cleaned.lower())
    if candidate in normalized_allowed:
        return True
    if normalized is not None and ACCOUNT_PLAN_EQUIVALENTS.get(normalized, frozenset()) & normalized_allowed:
        return True
    return False

(I.e. fall through to the lowercased raw string when normalization can't resolve it, instead of dropping the account.)

A quick regression test in tests/unit/test_plan_types.py would help -- something like:

def test_unknown_plan_passes_when_explicitly_allowed():
    assert account_plan_matches_allowed("future_plan", {"future_plan", "plus"})

def test_unknown_plan_blocked_when_not_in_allowed():
    assert not account_plan_matches_allowed("future_plan", {"plus"})

Once that's in I'll re-run @codex review and if it lands clean we can merge. The rest of the PR (capacity entries, equivalents map, openspec deltas) is fine.

Heads-up that I dropped a couple of follow-up notes on #557 too:

• The current prolite -> pro capacity values mean prolite accounts get the same routing weight as Pro, even though their actual quotas may be smaller. If routing telemetry shows prolite over-utilization later, calibrating capacity values down (or treating prolite as roughly plus * 5) would be a clean follow-up.
• sharenla's point chore(main): release 0.1.3 #4 (stale quota_exceeded / rate_limited not auto-recovering) looks like a separate issue from prolite -- the recovery logic in _compute_account_state already runs on every selection. If anyone has a concrete repro it deserves its own ticket.

Will re-trigger review after the regression fix.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5419de7772

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]

Allow unknown plans when explicitly listed

account_plan_matches_allowed now returns False whenever normalize_account_plan_type cannot resolve the account plan, which regresses the previous behavior of _filter_accounts_for_model (direct string membership). If the model registry starts emitting a new slug (for example a future plan type) and available_in_plans includes that same slug, those accounts will still be filtered out and selection can fail with no_accounts until code is updated. This should fall back to comparing the cleaned raw plan string against allowed_plans when normalization returns None so explicitly-allowed unknown plans continue to work.

Useful? React with 👍 / 👎.