Preserve exact fractional Unix timestamps in from_format

[FanWu-ai]

Pull Request Check List

• Added tests for changed code.
• Updated documentation for changed code.

Fix

from_format('0.000001', 'X') loses the microsecond; from_format('0.000012', 'X') raises ValueError; and from_format('-0.1', 'X') returns .100000 rather than .900000 in the preceding second. Negative x millisecond values are likewise incorrect. Float conversion also loses precision at large timestamps and can round the final valid microsecond beyond year 9999.

Parse the existing X/x grammars into exact integer microseconds and obtain calendar fields with stdlib epoch-plus-timedelta arithmetic. Timezone attachment remains unchanged. Add 26 regressions/controls and timestamp documentation examples.

Validation and CI correction

The first submitted head passed full local pure-Python and release-native checks, but hosted native testing failed on the retained year-1 case. CI uses Rust's debug profile; the existing native local_time helper panics with unsigned overflow there, while release arithmetic wraps. This exposed a gap in the initial local build-profile coverage.

The corrective commit avoids that helper using exact stdlib integer date arithmetic. Every test and documentation byte is retained, including year 1/year 9999. No Rust code, overflow settings, dependencies, skips or check configuration were changed.

Revised validation on Linux:

• Fresh upstream debug-native baselines on Python 3.10.22/3.12.14: 1,871 passed / 3 existing skips. The first submitted code independently reproduces the single debug-native failure on both versions.
• Final complete debug-native and pure suites on both Python versions: 1,897 passed / 3 existing skips each.
• Rebuilt release Python 3.12 wheel, native and pure complete suites: 1,897 passed / 3 existing skips each.
• Strict mypy (228 files), all six pinned hooks including Rust formatting/Clippy, strict MkDocs, debug/release packages, strict Twine and artifact-source comparisons pass.
• 40,412 exact-integer public-API checks and independent 4,127-case oracles pass. Explicit provenance probes verify debug overflow checks remain active while the corrected public formatter handles year 1.

The unchanged skips cover Windows-only behavior and upstream timezone-transition limits. Existing INFO-level docs anchor notices and Cargo packaging diagnostics remain. Replacement hosted CI and other platforms are not claimed passed until verified. Prepared and tested with AI assistance (OpenAI Codex); no human review is claimed.

[codspeed]

Merging this PR will not alter performance

⚠️ 1 benchmark spent significant time in system calls

System calls cannot be consistently instrumented, so they are not included in the measure, which understates the real cost. Please switch to the Walltime instrument to accurately measure system calls.

Measurement and system calls

• test_parse_iso8601

ℹ️ 1 benchmark has no flame graph

The flamegraph could not be generated because we failed to resolve the symbols in the profiling trace. This happens when we fail to harvest information about language frames.

• test_parse_iso8601

✅ 1 untouched benchmark

---

_{Comparing FanWu-ai:fix-exact-fractional-unix-timestamps-20261012 (d0b04f3) with master (7840d47)}

[Wkhanex]

I checked d0b04f3 in pure-Python mode on Python 3.12 using integer microseconds as the independent oracle. For 2,000 seeded signed offsets around the epoch, both X decimal-second strings and x millisecond strings matched datetime(1970, 1, 1) + timedelta(...) exactly (4,000 comparisons). This included negative subsecond values and offsets far enough from the epoch that a float would lose low-order digits. The last microsecond of year 9999 and the start of year 1 also remained exact. The from-format module passed 87 tests. This checks the formatter path only; I have not run the whole suite or native backend.

Review and local validation prepared with OpenAI Codex assistance.