perf(table): avoid cloning schema history for position_deletes partition IDs

[fallintoplace]

What changed

• Start synthetic position-delete partition field IDs at LastColumnID() + 1.
• Keep the existing reserved-ID checks and partition-field remapping.
• Avoid cloning and indexing every historical schema on each call.

Benchmark

• Machine: Apple M1 Pro, darwin/arm64
• Go: go1.26.3
• Setup: 64 fields per schema; median of 5 runs
• Command: go test ./table -run '^$' -bench '^BenchmarkPositionDeletesPartitionType/schemas=(1|16|128|1024)/fields=64$' -benchmem -count=5 -benchtime=1s

schemas	ns/op	B/op	allocs/op
1	40,476 -> 18,736	100,368 -> 48,960	94 -> 56
16	347,383 -> 18,448	879,377 -> 48,960	570 -> 56
128	2,695,818 -> 18,115	6,692,085 -> 48,960	4,101 -> 56
1,024	20,795,189 -> 16,137	53,193,992 -> 48,960	32,328 -> 56

The optimized path is effectively flat as schema history grows.

Tests

• Added coverage that the allocator does not request schema history.
• Added coverage for skipping reserved IDs after LastColumnID().

[zeroshade]

The history-independent allocation is a substantial improvement, but the new starting-point calculation can produce duplicate or out-of-range field IDs.

Blocking — Guard synthetic field-ID allocation (table/inspect_position_deletes.go:817–831)

metadata.LastColumnID()+1 is used without validating it against the current schema or the signed 32-bit field-ID boundary.

I reproduced two regressions:

1. The metadata parser accepts a schema field with ID 2 and last-column-id: 1. This change assigns the partition field ID 2, after which PositionDeletesSchema panics:

   invalid schema: multiple fields for id 2: x and x
   

   The merge-base implementation assigns an unused ID and succeeds.

2. With LastColumnID() == iceberg.MaxStructFieldID and 193 active partition fields, the final generated ID is 2147483648, outside Iceberg’s signed 32-bit field-ID range. This also wraps on 32-bit platforms.

Please keep the fast path but seed collision checks from the current schema and make allocation overflow-safe. If the range above LastColumnID is exhausted, fall back to low unused IDs or return a structured error. Regression coverage should include both cases above.

---

This review was drafted by an AI-assisted tool and confirmed by an Iceberg Go maintainer. After you've addressed the points above and pushed an update, an Iceberg Go maintainer — a real person — will take the next look at the PR. The findings cite the project's review criteria; if you think one of them is mis-applied, please reply on the PR and a maintainer will weigh in.

More on how Iceberg Go handles maintainer review:
CONTRIBUTING.md.