[major] Export codegen emits select_option() with no arguments — recorded option value silently dropped (false-fails and false-passes)

[tlemmonds]

Summary

testmd export (Playwright/Python) generates select_option() with no arguments for a recorded dropdown selection, even though the recording carries the selected value. The generated line selects nothing, and the test silently derails from that point on.

Evidence

CLI 0.6.4, 2026-07-24. Recorded action (from actions.ndjson — note action_params.value is present and the selector is a clean testid):

{"action_type": "select",
 "action_instruction": "Selecting the sample applicant persona dropdown option",
 "selectors": [{"selector": "internal:testid=[data-testid=\"persona-select\"s]", "score": 100}],
 "action_params": {"value": "Dana Ortiz · Decline", "label": null, "index": null}}

Generated code for that action:

element_1 = page.locator("internal:testid=[data-testid=\"persona-select\"s]")
await element_1.select_option()   # <-- no value/label/index — selects nothing

Suspect: the non-ASCII · (U+00B7) in the option value trips the argument serialization.

Impact — silent cascade, both false-fails AND false-passes

In our lending AUT the dropdown loads an applicant persona (prefills every form field). With the selection dropped:

• persona never loads → all fields stay empty → bank verification can't complete → the decision engine's first rule returns RFAI
• a test expecting Decline fails with RFAI (reproduced on TMS cloud execution AND a local run of the same export — deterministic)
• worse: tests that expect RFAI false-pass on the empty form — green results that validated nothing

Compounding issue (same export)

The decision extraction in the same generated file is re-derived as a brittle text-glob with no expected value:

locators=['internal:text="DecisionDeclineYour"i'], expected_value=''

even though the freshly authored recording carries internal:testid=[data-testid="decision-outcome"s] with expected_value: "Decline". Codegen discards the recorded anchors it should trust — same family as #150.

Ask

1. Emit the recorded value/label/index in select_option(...) — and fail codegen loudly if none survives serialization, rather than emitting a no-op call.
2. Add a codegen test with non-ASCII characters in option values.
3. Prefer recorded testid locators and recorded expected values over re-derived text globs (see [major] Exported code has no wait before post-action extractions — author-green tests fail deterministically on fast executors #150).

Workaround

We added a /apply?persona=<id> deep-link testability hook to the application so TMS-executed tests avoid the <select> entirely — workable for us, not for teams who can't modify the AUT.