[1932819331hy-debug]

Summary

 Design-clock period counts near a floating-point boundary
- trust floor(elapsed / rounded period duration)
+ check the adjacent deadlines using the existing intendedEnd calculation

Fix both sides of an ordinary fractional-rate boundary:

• At 200 ms/block, five 1003-block periods take 1003 seconds. The report previously counted only four expected periods while reporting zero drift.
• At 150 ms/block and three blocks/period, a schedule starting at second 1 has its hundredth deadline at second 46. The immediately preceding representable timestamp previously counted that period too early.

Check the next and already-counted deadlines without an epsilon or loop. Keep the existing intendedEnd arithmetic, input rules, RPC behavior and other report fields unchanged.

Evidence

• Before: the same 22 new native tests have 11 passes / 11 assertion failures on main 13184ff111532c172f37d54d59f00e7eea6b9a23.
  After: all 22 pass; native npm test passes 45/45.
• Independent AI-assisted review passes 44/44 tests, including both public entry points, the timestamp immediately before/at/after a deadline, fractional and zero origins, caps and unchanged fields. Eight deliberately broken variants are detected.
• Local composition with fix: reject unsafe schedule block counts before projection #1 and fix: reject primitive RPC responses without retrying #2 passes 72/72 in both application orders, producing the same tree. Their branches and scopes are unchanged.
• The independent review ran 32 commands. Its 12 nonzero outcomes are retained and classified: two baseline reproductions, eight negative variants, the offline-guard self-test, and the decimal diagnostic below.

Numerical limit: the broader decimal-millisecond diagnostic still fails with 768 discrepancies among 37,800 cases at low/zero origins. It compares independently rounded decimal totals with the library's existing binary operation order. This patch preserves the observable intendedEnd convention. The separate reported-deadline consistency sample passes, but this is not a full-domain precision claim.

All execution used Node 24.22.1 and synthetic in-memory transport. Node 18/20/22, other operating systems, live RPC, actual CLI/Action execution and hosted CI remain unverified. Build, lint and typecheck scripts are not configured.

AI assistance was used for implementation, testing and independent review. This is a source-backed contract repair, not completion of a separate maintainer issue.

Merge Danger

Door: Two-way. A localized arithmetic correction with no migration.

Blast Radius: Reporting.

Expected-period and periods-ahead counts can change by one at the demonstrated boundaries. Deadline values and numeric policy are unchanged.