[hanzhe-one]

Purpose of the pull request

Closed: #1303

ExcelDateFormatDetector cached date-format detection in a thread-local Map<Short, Boolean> keyed by the format index alone. Custom format indexes (>= 164) are defined per workbook, so the same index can map to different format strings. Once an index was cached, later queries with a different string returned the earlier result:

isADateFormat((short) 164, "yyyy-mm-dd") -> true
isADateFormat((short) 164, "0.00")       -> true   // should be false

Reverse order made a date format report false. isADateFormatUncached returns the correct values, so the cache disagreed with the detector.

What's changed?

• The thread-local cache is now Map<String, Boolean> keyed by formatIndex plus the format string; a null format string maps to a stable key.
• Detection logic, public signatures and cache clearing are unchanged.
• Existing DateUtilsTest.test_isADateFormat_Cache now asserts the new key shape instead of reflecting on a Short key.
• Four new tests cover: same index with different strings in both orders, null format string, and the internal-date-index path.

The cache is cleared in WriteContextImpl.finish and ExcelAnalyserImpl.close, so a completed read or write normally starts fresh. This change removes the reliance on that clearing for correctness when the same thread queries one index with different format strings.

Validation

• New tests fail against the unfixed implementation (3 failures) and pass with the change.
• ./mvnw -pl fesod-sheet test: 1,793 tests, 0 failures. DateUtilsTest: 94 tests, 0 failures.
• ./mvnw spotless:check: passes.

The only errors in that run are 10 in UrlImageConverterTest, all IOException: Unable to establish loopback connection when the test starts an HttpServer. The same 10 errors occur on unmodified main in this environment, so they are a sandbox networking restriction rather than a regression from this change; CI runs without that restriction.

Note: open PR #1260 also modifies ExcelDateFormatDetector when centralizing thread-local cleanup; whichever lands second needs a trivial rebase.

Checklist

• I have read the Contributor Guide.
• I have written the necessary doc or comment.
• I have added the necessary unit tests and all cases have passed.

[Mikkey-f]

Thanks @hanzhe-one — I reproduced the bug and verified the fix locally (Windows 11, JDK 17).

Root cause: DATE_THREAD_LOCAL is keyed by the format index alone. Custom indexes are workbook-scoped, so the same index can map to different format strings. Calling DateUtils.isADateFormat(164, "yyyy-mm-dd") then (164, "0.00") returns true twice; the reverse order returns false twice, while isADateFormatUncached gives the correct pair.

Verification

• Red: with only ExcelDateFormatDetector reverted (your tests kept), 3 new tests fail as described — both orders plus the null format string; the internal-index test passes either way.
• Green: DateUtilsTest 94/94.
• Full suite: 60 + 1,793 tests, 0 failures; spotless:check clean. I didn't hit the UrlImageConverterTest loopback errors here, consistent with your sandbox note.

Design check: Mapping a null format string to "" is behavior-preserving — isADateFormatUncached treats both as empty after the internal-index check — and the NUL (0x00) separator keeps the composite key unambiguous. The cache lifecycle is unchanged: removeThreadLocalCache() still runs from ExcelAnalyserImpl.close and WriteContextImpl.finish, so normal read/write paths remain operation-scoped.

One nit (non-blocking): test_isADateFormat_Cache now asserts size and containsValue(true), but the key shape is the point of the fix — asserting the exact composite key (formatIndex + NUL + formatString) would pin it directly.

LGTM — the cache now agrees with the uncached detector, and both orders are covered.

[Copilot]

🟢 Approval recommended

No active review issues remain for the date-format cache-key change.

No open issues 🎉

What changed in this PR

Fixes #1303 by making date-format cache results depend on both the format index and format string, preventing stale results when workbooks reuse an index for different formats.

Changes:

• Uses a composite string cache key with stable handling of null format strings.
• Updates cache assertions and adds regression tests for differing strings, call order, null strings, and built-in date indexes.
• Preserves detection logic, public signatures, and cache cleanup behavior.

File	Description
fesod-sheet/​src/​test/​java/​org/​apache/​fesod/​sheet/​util/​DateUtilsTest.java	Updates cache assertions and adds regression coverage.
fesod-sheet/​src/​main/​java/​org/​apache/​fesod/​sheet/​util/​ExcelDateFormatDetector.java	Keys cached results by format index and string.

_{Review effort: Balanced}