Three additional open-issue pairs — registered follow-up

Registered before these solver and review calls. Three previously unattempted open issues from three repositories have complete, nonempty public discussions and different retrieved comment-ID sets for the two arms. The manifest, SHA-256 c9daf5df6345bdd6e1e3f57de62ad954208b2871a58144d578d8a6320c842a1d, freezes source commits, contexts, partial checks, paired order, models, budgets and stop rules. This selection follows earlier outcomes and occurred while batch 30 was running. These projects also appear in earlier batches; distinct issue IDs do not establish statistical independence.

Open issue	Frozen partial acceptance	Proposed reference
Day.js #2442	Japanese morning and afternoon strings parse to the expected hours; selected English AM/PM cases remain valid.	Source-only PR #2983. The checks do not cover every locale or meridiem ambiguity.
validator.js #2036	Selected invisible characters in an email local part are rejected by default; ordinary ASCII and a selected UTF-8 address remain accepted.	Source-only PR #2648 with its default manually inverted because the original allows these characters by default. Broader compatibility and all Unicode categories are not checked.
math.js #3604	Selected negative, noninteger inputs return complex LogGamma values from the later issue discussion; positive input and zero retain selected behavior.	Source-only PR #3648. The original issue asked about a real value; choosing the complex branch follows subsequent discussion and is not evidence of an accepted final specification.

The qualification ledger preserves the expected baseline failures, reference passes, regression controls and both original and adapted references. These checks are narrow and proposed references do not establish maintainer acceptance. Conditions, AI reviews and adjudication, bounded 503-only retries, sandbox and failure policy follow batch 30. The estimated budget is below $1, not a hard cap. Public issue text or discussions may reveal solutions or be present in model training; distinct model labels use the same proxy with unverified provenance. Short pre-seeded histories are not a production session or a test of autonomous memory management.