macOS workflow-expression timeout regression loses its expected result in full CI

[RolandUI]

Full CI 36057932181 on exact commit 2db6dc6 fails the macOS Runtime Test job: RuntimeExpressionTests.WorkflowWaitsReuseExpressionsResolveVariablesAndRetainTimeoutOperands throws NullReferenceException at RuntimeExpressionTests.cs:132 after ~1 second. Aggregate: 808 passed, 1 failed, 5 gated native skips. Windows full suite, all three native QA jobs and Linux Installer pass. Retained log: artifacts/agent-qa/expanded-campaign/ci-2db6-failed.log; original CI failure must remain visible.

Investigate the missing workflow/expression result at the assertion boundary; distinguish a test timing/precondition assumption from a product error. This is a confirmed failing check, not an established root cause and not evidence that #176 caused it.

Acceptance:

• Preserve the original failure and add useful result/diagnostic assertions instead of dereferencing an absent value.
• Reproduce or narrow the condition with the actual workflow path; fix the supported cause without weakening timeout semantics or blindly increasing product deadlines.
• Run focused regressions and a fresh full macOS/Windows gate, recording any remaining uncertainty explicitly.

Queued behind #178; #176 table-specific Direct/standalone acceptance passes but the comprehensive gate remains failed. No release/publication.

[RolandUI]

Read-only narrowing while #178 remains active: the failing line dereferences timeout.Value.Steps[0].WaitObservation.Expression. SemanticWorkflowRunner.WaitForConditionAsync replaces lastObservation with MissingObservation when an expression poll returns a failed CoreResult (around lines 1944-1955); an OperationCanceledException instead preserves the previous observation. Investigate whether a deadline-boundary failed result overwrites previously captured operands, or whether this run never completed a poll. The original log cannot distinguish these cases. Do not claim a cause or modify deadlines until a controlled regression and full result diagnostics resolve it.

[RolandUI]

Starting #179 as the sole implementation issue while #178 waits in review for its expanded native CI gate. Preserve the original macOS failure, add diagnostic assertions, and exercise the deadline/last-observation boundary through the real workflow and controlled bridge responses. Determine whether unavailable/late polls overwrite useful expression operands or the test assumes a poll completes within 500 ms. Keep product timeout semantics; validate focused regressions and fresh complete Windows/macOS tests. Do not relabel a passing unchanged rerun as a fix.

[RolandUI]

Controlled real named-pipe deadline regressions now reproduce two distinct failures before the fix (expression-boundary-before-results/expression-boundary-before.trx: 2 failed, 1 passed). A completed false evaluation followed by a protocol error is relabeled missing and loses its expression; a first poll that never responds returns a generic timeout without an explicit unavailable observation. A completed false evaluation followed only by a stalled poll already retains its evidence correctly. The implementation keeps historical operands/timestamps on failed polls and explicitly identifies a deadline with no completed evaluation. Product deadlines are unchanged. The former 500 ms UI integration assertion is split into real variable-resolution/false-assertion coverage plus controlled three-second deadline cases, without claiming which poll sequence occurred in the original macOS run. Focused validation is running. The unchanged 2a80ba1 full macOS test stage passed; that alone is not a fix.

[RolandUI]

e2f436d is committed and pushed. Final focused validation passes 18/18 expression evaluator, controlled named-pipe deadline and real UI/CLI/MCP tests (expression-final-results/expression-final.trx, 35 seconds, no warnings/errors). The old subsecond UI wait was split into supported real wait/branch substitution checks and three controlled deadline cases. A diagnostic rerun caught an invalid test-authored assert_state request; this was corrected to the supported expression branch, not treated as a product defect. One earlier setup-only frame-readiness failure remains in expression-after; setup now includes the actual response in any recurrence. Full CI is dispatched with the expanded 40-minute native-lab budget. Original 36057932181 macOS poll sequence remains unknowable from its old NRE log; the controlled evidence loss is reproduced and fixed. Moving to review while the full gate runs; #180 is the next implementation issue. No release.