perf: galloping search in sealed LID iterators' NextGeq

[radmirnovii]

Description

NextGeq in the sealed LID iterators runs a binary search over the remaining
chunk slice on every call. The AND/OR zigzag sends monotonically increasing
targets while the slice narrows from the consumption edge, so the answer is
almost always near the edge — instrumented on a 2M-doc fraction: ~50% of calls
land at distance 0 and 99.9% within 16, while binary search pays
log2(remainder) ≈ 14 probes regardless. This PR replaces it with an
exponential (galloping) search from the edge, ~2·log2(d) probes — the same
uncapped scheme as Lucene's
IntArrayDocIdSet.advance
and RoaringBitmap's
advanceUntil,
which seq-db already depends on for skip masks.

The worst case is bounded at ~2× binary (2·log2(d) ≤ 2·log2(n)) and far jumps
are self-limiting (jumps over one chunk sum to at most its length): an AND
built deliberately from a skewed pair (freq ratio ≈1450, 71% of calls in the
far zone) measures at parity. A capped variant was measured and rejected — it
hurts the common zone and rescues regimes that did not occur across ~6M
instrumented calls (max distance < 128).

Benchstat, n=10, in-process Sealed.Search, 2M-doc fraction, vs main:

scenario	diff
3-token AND	−21.8%
(a OR b) AND c	−10.8%
2-token AND	−4.5%
skewed ANDs (ratio ≈285 / ≈1450)	~
13 other scenarios (needle hit/miss, aggregations, histogram, wildcard/contains, fetch, active fraction, pure OR, AND NOT)	~

Search results are byte-identical to main on a 57-query corpus (IDs and
order in both sort directions, aggregation bins, fetched docs).

Note: fracmanager/TestCapacityExceeded is red on current main (a73114c)
independently of this change.

---

• I have read and followed all requirements in CONTRIBUTING.md;
• I used LLM/AI assistance to make this pull request;

Model: Claude Fable 5 (claude-fable-5), via Claude Code
Prompt: interactive session, no single prompt. Task: performance research of
seq-db search internals (profiling, microbenchmarks, adopt/reject verdicts
backed by numbers). Directives shaping this PR: deep-dive galloping; prove
edge-locality with in-engine instrumentation rather than synthetic tests;
write a standalone PR with edge-case/regression justification and prior-art links
Extent: code (gallop.go, test, call sites), benchmarks, and this PR text
were AI-generated under the author's direction and review.

[eguguchkin]

@seqbenchbot up main

[seqbenchbot]

Oh-oh, @eguguchkin ('>o.o)'>!

Something went wrong and I couldn't process your request.
Please take a closer look at error message:

unexpected argument up

[eguguchkin]

@seqbenchbot --help

[seqbenchbot]

Hey, @eguguchkin *\ (^_^)}!

Below is the help message you've requested.
Please notice that I support shell-like interface for interaction!

Usage: @seqbenchbot <command>

Flags:
  -h, --help    Show context-sensitive help.

Commands:
  start bulk <reference> [flags]
    Bulk write scenario

  start mixed <reference> [flags]
    Mixed scenario

  start search-logbench <reference> [flags]
    A collection of different queries, each cold and hot

  start search-regular <reference> <query> [flags]
    Benchmark a single query, cold or hot, for a duration

  start search-aggregation <reference> <func> <fields> ... [flags]
    Benchmark a single aggregation, cold or hot, for a duration

  stop <id>
    Stop a running benchmark early

  compare <baseline> <comparison>
    Compare two finished runs

  list
    List benchmark runs

Run "@seqbenchbot <command> --help" for more information on a command.

[eguguchkin]

@seqbenchbot start search-aggregation

[seqbenchbot]

Oh-oh, @eguguchkin ('>o.o)'>!

Something went wrong and I couldn't process your request.
Please take a closer look at error message:

expected "<reference> <func> <fields> ..."