[Bug]: two shipped public smokes now contradict #7587's projection contract and fail on main (hot-path quota payload) #7653

Preflight

• I searched existing issues and discussions for this behavior.
• This report is not an unpatched security vulnerability.
• The report and attachments contain no credentials or private/internal material.

Issue origin

Reproduced in a real environment (native Windows source checkout).

LoopX version or commit

9dcdc0bacd455a8cf6f6146f052ffc4df0f5d04a (current main, LoopX 2.2.0).

Host or runtime surface

Not host-specific; reproduced on native Windows 11 / CPython 3.14.

LoopX area

Control plane (goals, todos, quota, scheduler, registry, runtime)

Problem

Two shipped public smokes assert a hot-path shape that the current projection no longer produces, so both fail on a clean main checkout. The underlying cause is a narrower version of the behavior #7587 intentionally introduced.

loopx/control_plane/quota/diagnostic_projection.ts classifies each diagnostic row with knownItem():

const scalar = (value: unknown) => value === null || ["string", "number", "boolean"].includes(typeof value);
return item !== null && (...) && Object.entries(item).every(([key, field]) => ITEM_FIELDS.has(key) &&
  (scalar(field) || (Array.isArray(field) && field.every(scalar))));

ITEM_FIELDS is TODO_CANONICAL_READ_RECORD_FIELDS, and that contract declares structured fields — resume_condition, decision_scope, global_gate, required_decision_scopes, blocked_retry, completion_continuation. A row carrying any of them therefore fails the scalar test, is classified unfamiliar, and is preserved verbatim by #7587's rule. #7587 named exactly this case in its own description: "a task carrying a structured recovery condition".

Measured effect (bounded, not unbounded). The per-lane ceiling still applies: deferred_items keeps at most deferred_visibility_limit = 8 rows. What changed is that lanes previously omitted entirely are now emitted, and the retained rows keep a duplicated title and the full nested condition. Through the real project_asset_todo_summary builder plus the hot-path projection on 9dcdc0bac:

That is a 6.0x (10,004-byte) increase for the ordinary shape of a Goal with parked work, re-sent on every automatic turn.

Minimal reproduction

From a clean source checkout of 9dcdc0bac:

python3 examples/control_plane/todo-durability-fixture-smoke.py
python3 examples/control_plane/control-plane-integrated-canary-smoke.py --max-seconds 400

Observed at 9dcdc0bac: the first exits 1 at assert "deferred_items" not in agent_todos; the second fails at assert "title" not in item. Both exit 0 at 97d38f9a8, immediately before #7557.

Expected behavior

This issue does not assert which side is correct; both are defensible and the maintainer decides:

• If fix(quota): bound mixed diagnostic summaries without losing unfamiliar records #7587's preservation rule is authoritative, the two smokes encode the retired contract and should be updated to the accepted shape.
• If the hot path must still omit these lanes, knownItem() should treat contract-declared structured fields as known rather than unfamiliar — the predicate currently conflates "field value is structured" with "field is not part of this build's contract".

Either way, main should not ship two smokes that disagree with the merged projection contract.

Why this matters

1. The nightly full-public sweep has been red since 2026-10-28 (ccf6aacc, smoke-fleet-health failure) after passing on 2026-10-27 (98c4ba31); these two smokes are part of that sweep.
2. The disagreement is silent: the projection reports omitted_lanes without these lanes, so a reader cannot tell the shape changed.
3. The cost recurs per turn on the absolute_hot_path surface rather than once.

Additional context

• Regression boundary: 97d38f9a8 passes; 4b2b97b7b (Preserve complete context in shared quota diagnostics #7557, which added knownItem) fails.
• fix(quota): bound mixed diagnostic summaries without losing unfamiliar records #7587 then scoped the fallback to individual rows and locked the current behavior in tests/control_plane/test_quota_cli_projection.py (the resume/future/schema/scalar parametrization), so any change here must also move those tests.
• Corrections to my own earlier draft: growth is bounded per lane, not unbounded, and I incorrectly reported the kernel-contracts step as failing — it passes on 9dcdc0bac. See the correction comment below.