v1.1.1

• Fix toMillis() violating its exact-ceil contract above 2^53 microseconds (~285 years): float division lost integer precision and the result was off by ±1 ms in both directions (including round-DOWN, the exact failure the contract forbids). Now pure integer arithmetic (intdiv + remainder), exact over the full range.
• Fix millis() — the one factory without an overflow guard: past PHP_INT_MAX the product silently became a float and the constructor threw an unrelated TypeError instead of the contractual InvalidArgumentException('Duration overflow: ...').
• Fix integer inputs to seconds()/minutes()/hours()/days() losing precision through the float path once the product exceeded 2^53 (e.g. seconds(9_223_372_036_853) came out 192 µs off): integer inputs now stay in integer arithmetic end to end, with the same overflow contract as plus().
• Sidestep the PHP >= 8.4 round() regression that returns n+1 for even integer-valued floats in [2^52, 2^53) (corrupted e.g. days(86165) by +1 µs on 8.4/8.5): round() is now skipped for float products at magnitude >= 2^52, where every representable double is already integral and rounding is a mathematical no-op.
• Docs: rewrite the __toString() docblock — the unit is chosen by boundary fall-through of rounded values, not by a "displayed value >= 1" rule (fall-through can legitimately print "0.999999min"); note the deliberate half-to-even vs half-away tie asymmetry between %g and the plain-integer day rendering. Both sub-unit fall-through outputs and the half-away rounding of the float factories are now pinned by tests.
• Test hardening: toMillis property against an integer-ceil oracle over the full [0, PHP_INT_MAX] range; microsRoundTrip/millisRoundTrip generators widened to the full valid ranges (the old caps were orders of magnitude below the bug zones); exactness properties and examples for the integer factories.