Four issue pairs — registered follow-up

Registered before these solver and review calls. Four previously unattempted open issues from four repositories have complete nonempty comment snapshots and different retrieved comment-ID sets in the two arms. The manifest, SHA-256 af0ed528d3d366e8fff4674b68b85cc5871bd8a735934130935059550909b4a9, freezes source commits, partial checks, contexts, paired order, models, budgets and failure policy. Selection follows earlier outcomes; distinct issue IDs and repositories within this batch do not prove statistical independence. The math.js task and a batch 27 issue touch the same source file.

Open issue	Frozen partial acceptance	Proposed reference
qs #432	With allowEmptyKeys: true, repeated empty keys are preserved alongside ordinary keys; blank delimiters do not create keys and default behavior remains unchanged.	Narrow manual port of PR #433, whose original source did not apply to the pinned parser.
Express #7350	A missing view ending in a dot reaches the callback with the normal lookup error; ordinary missing and successful views still use their callbacks.	Source-only PR #7392.
Cheerio #581	HTML attribute getters accept uppercase lookup names while XML getters remain case-sensitive; ordinary lowercase and absent attributes remain valid.	Getter-only manual port of PR #3306, whose original source did not apply to the pinned version. Setters and the data API are outside this partial check.
math.js #3620	unit('m ^ (1)') and a second parenthesized exponent parse like their unparenthesized forms, malformed parentheses fail, and ordinary unit parsing remains valid.	Source-only PR #3681; the discussion does not establish that maintainers have accepted this behavior.

The qualification ledger records baseline failures, reference passes and controls on both, plus the original and adapted reference patches. The initial Express success control asserted before its asynchronous callback and failed on both versions; it was corrected and rerun before registration. A preliminary qs baseline require was also run in the dedicated VM outside the inner network-disabled sandbox; the recorded qualification runs used the sandbox. Dependencies were installed without lifecycle scripts, with locks recorded in the manifest. Checks are narrow and proposed references are not maintainer acceptance.

After four batch 27 attempts stopped when the model requested more than four tools in one response, this batch changes the runner: it executes at most the first four, records each remaining call as not executed, and returns that result to the model so it may continue within the same turn budget. The change applies equally to both arms and is tested, but makes cross-batch comparisons exploratory. Model names, AI blind-rating rubric, separate adjudicator and bounded 503-only retries follow batch 27. The estimated budget is below $1, not a hard cap. Shared gateway provenance, solution exposure, partial checks and short issue histories prevent production or independence claims.