bindings docs: known issues for upstream #9014, #9017, #9021, and #9027, with workaround tests and strict xfail tripwires

[tae898]

Closes #106.

Four engine findings from today's schema hunt reach Python users and are wrong answers or wrong stored values, which is what this page lists. Each was measured through the bindings on a 26.10.1 snapshot, with the workaround checked on the same reproduction.

• CREATE PROPERTY ... (mandatory true, notnull true) is accepted over records that lack the property, though ALTER PROPERTY refuses the same constraint since #8956, and SQL ORDER BY over an index on it then drops those records; openCypher CREATE CONSTRAINT and ALTER PROPERTY MIN, MAX, and REGEXP skip the check too ArcadeData/arcadedb#9017: CREATE PROPERTY ... (mandatory true, notnull true) over records that lack the property, then ORDER BY over an index drops those records (new on main since 6212446 and 35eb31e). Workaround: give every record the property first, or relax the constraints with ALTER PROPERTY; both returned all five records.
• An index on an INTEGER or LONG answers for a bound truncated to the key type: SQL WHERE i = 12.5 returns the record holding 12, i >= 12.5 returns 12, i < 12.5 misses 12, and l = 1e19 returns Long.MAX_VALUE; openCypher n.i < 12.5 misses 12 since 179cf8dcdc ArcadeData/arcadedb#9021: an index on an INTEGER or LONG answers for a bound with a fraction as if it were rounded (i = 12.5 returns 12, i < 12.5 misses 12, a float of 1e19 finds Long.MAX_VALUE). Workaround: round the bound in the direction of the operator; the rounded bound matched the unindexed scan for all four range operators.
• A value that cannot be converted to a declared numeric type is stored as NULL and the write reports success: true, [1, 2], {a: 1}, and a LocalDate written to an INTEGER property all read back null (the string 'abc' is refused) ArcadeData/arcadedb#9014 and Type conversion stores '' as 0 in SHORT, INTEGER, LONG, FLOAT, and DOUBLE properties (BYTE and DECIMAL refuse it), and reads a number as a BOOLEAN through intValue(), so 0.5 and 4294967296 are stored as false and WHERE flag = 0.5 matches the false records ArcadeData/arcadedb#9027: a bool, list, or dict written to a declared numeric property is stored as NULL, and '' as 0 in SHORT, INTEGER, LONG, FLOAT, and DOUBLE. Workaround: convert in Python first.

tests/test_declared_type_known_issues.py has a test of each workaround and a strict xfail per finding that fails the suite when an engine fix reaches the wheel. I ran the xfail tests with --runxfail to confirm each fails on its intended assertion ([1, 2, 3, 4] against five records, [12] against [], DID NOT RAISE, 0 != 0). The page header now says what a Tests: line means. ArcadeData#9024 has no entry: a Python int past int64 raises OverflowError before the engine sees it.

Rebased onto #105 (the two entries append to the same place; both kept). Full suite on a wheel built from the cecb369 engine: 503 passed, 3 skipped, 11 xfailed. black and bandit (run after the formatter this time) are clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DeKEhHXJTy1H1GgbfBZ8mL