Five distinct-project issue pairs — registered follow-up

Registered before these solver and review calls. These five currently open issues were not used in the earlier instrumentation batches. Each has a complete nonempty comment snapshot; the keyword and Memex contexts retrieve different comment-ID sets. The snapshots, source commits, checks, randomized paired order, budgets and runner are bound by the manifest, SHA-256 d14c3a849a7583fcefd23bf4169f57046638bd2e0d95b9175616f4a79a930692. The five repositories are distinct within this batch, but issue identities alone do not establish statistical independence.

Open issue	Frozen partial acceptance	Proposed source-only reference
validator.js #2430	Scientific notation examples including 2e+21 and -3.5E-8 are accepted; malformed exponents are rejected, and ordinary numeric and no_symbols controls remain valid.	PR #2447
Ramda #2137	assocPath preserves own symbol properties on copied outer and nested objects without mutating the input; ordinary nested updates remain valid.	PR #3506
date-fns #4096	French abbreviated month inputs jan, fev and aout parse to the expected months; existing accented and invalid-input controls remain valid.	PR #4119
math.js #3134	Comparing 9 °C converted to °F, K or °R against 10 °C preserves ordering, equivalent temperatures compare equal, and ordinary length controls remain valid.	PR #3705
Day.js #2281	Lowercase localized formats work for a fixed locale without its own format table; English uppercase and plain-date controls remain valid. The locale is a controlled variant of the reported missing-format case.	PR #3217

The qualification ledger records assertion failures on each pinned baseline, passes on each source-only reference, and controls passing on both under the network-disabled VM sandbox. The checks are narrow, proposed PRs are not maintainer acceptance, and the date-fns source is loaded with a test-only TypeScript hook. The issue discussions and public PRs can disclose proposed solutions, and can appear in model training.

Each pair uses a randomized arm order and the same bounded search/read/exact-edit/public-test tools. gpt-6-astra is the solver; deepseek-flash and qwen3.8-flash are blinded reviewers, with gpt-6.1-sol adjudicating valid score disagreements. These are distinct model names behind one proxy, not verified independent model providers or human assessors. The adjudicator was selected after a separate small route probe succeeded, which does not establish future availability. Only explicit HTTP 503 responses may receive at most two bounded retries; other failures and timeouts stop the stage and are retained. The target is below $1 for this batch, an estimate rather than a verified hard cap. Short seeded issue discussions do not establish long-running production effects or autonomous memory management.