Support integer, numeric, double and date CDX index keys

[chrisrichards]

Closes #292.

Summary

Non-character index tags now work as automatic access paths. In the foxprodb fixtures this unlocks the integer-keyed primary/foreign-key tags — where CONTACT_ID = 1 is now an index seek:

select CALL_ID from calls.dbf where CONTACT_ID = 1      → index seek (=) on tag 'CONTACT_ID'
select CALL_ID from calls.dbf where CALL_ID between 4 and 9 → index range scan (between) on tag 'CALL_ID'
select * from calls.dbf order by CONTACT_ID             → index order scan on tag 'CONTACT_ID'

The key transforms (CdxKeyEncoder)

• Integer (I): 4-byte big-endian with the sign bit flipped. Verified against the fixtures byte-for-byte — before writing any code, a diagnostic dump confirmed 1 → 80-00-00-01 on real VFP data, and a permanent test now decodes all 32 keys in both calls.CDX tags and matches them to the column values of the records they point at.
• Numeric/Float/Double (N/F/B) and Date (D): 8-byte big-endian IEEE 754 double — sign bit flipped for non-negatives, all bits flipped for negatives (the standard total-order transform, per the same xBase format references the DBF reader already cites); dates as Julian day numbers using the identical epoch constant as DbfValueDateTime (anchors tested: 1970-01-01 → 2440588, 2000-01-01 → 2451545). No fixture has an N/D-keyed tag, so these are covered by exact-bytes and order-preservation unit tests rather than differential ones — contributions of VFP files with such tags welcome.
• DateTime (T) and Currency (Y) keys remain unsupported (fractional-day and int64-vs-double encodings unverified); tags on those columns fall back to scans.

Planner semantics worth reviewing

• Integer bounds are adjusted to the integer domain: CALL_ID >= 9.5 seeks from 10, = 1.5 and between 8.5 and 8.9 are provably empty without touching the table, and bounds beyond int range either return nothing or defer to a scan.
• Double/date strict bounds stay inclusive at the key level: converting a decimal literal to a double can collapse a strict boundary onto the bound itself, so >/< include the boundary in the index range and let the always-applied residual filter drop the boundary rows — the superset invariant does the safety work.
• Binary keys pad with 0x00, not the character space pad — used both in comparators and duplicate-run stabilization (integer FK runs like CONTACT_ID = 1 now exercise that path with real duplicates).
• Key-width sanity checks per type (4 for I, 8 for the double family) guard against mismatched tags.

Test plan

33 new tests: encoder exact-bytes/round-trip/monotonicity (integers, doubles, dates, Julian anchors), the 32-key fixture round-trip proof, and a differential suite over calls.dbf (equality incl. duplicate runs/no-match/non-integral/out-of-range, all four range operators incl. flipped operands and fractional bounds, BETWEEN incl. the empty fractional window, integer parameters, residual predicates, ORDER BY satisfied by seeks and by a duplicated-key index scan — exercising stable duplicate ordering — and the typed-builder path). One phase-6 test that asserted integer tags fall back was removed, superseded by the positive coverage. Full suite: 415 passed, 1 pre-existing skip, zero warnings on both TFMs.

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud