[GordonYuanyc]

Adds --trials T, --trial-start S and --trial-mode fresh|relabel|reseed to approxbench sketchbench, so many independent accuracy trials run in one process instead of one process per trial. Trial t writes the same record whatever S and T are. Each record carries bench.trial and bench.trial_mode, the per-trial sketch config and the per-trial dataset descriptor. --flat writes one line per trial. --per-group-out gains a trial column, and all trials go to one file. A failed trial names the trial and its derived seeds. A run asking only for accuracy now skips the time-based warmup.

• fresh: every distribution seed and the burst seed become mix(seed, t) for t >= 1, and the key columns are offset.
• relabel: the table is drawn once and only the key columns are offset (integers and f64 + t*2^40, strings prefixed "t:"). The membership comparator's absent probes are offset too.
• reseed: the table is drawn once and the row's seed field becomes mix(base, t). Only rows whose wrapper uses the seed accept it: kll-percall, kll-cdf, kll-dynamic, hydra-kll, hydra-kll-cdf, coco, uniform-sampling, nitro-count-min, nitro-count-sketch (all lib).
• datagen: new ColumnSpec.key_offset (a number, or a prefix for strings), applied after generation and serialized only when set. mix_connected_bits is renamed splitmix64 and made public so mix reuses it.

Decisions:

• f64 key columns get the t*2^40 offset too. A key the offset does not move exactly (any non-integer f64 key) is refused, naming the column.
• No truth mapping under relabel. The table is generated once, then cloned and offset per trial, and rows::measurements recomputes the truth. reseed reuses the table the same way.
• The per-trial offset is stored in the descriptor's key_offset, so --trials 1 --spec <descriptor> reproduces a trial.
• mix() is documented in the --trial-mode help text.
• No --trial-threads.

Acceptance:

1. 94 rows in --list-impls. 3 measure no accuracy. I compared --trials 1 against main f31707d on each row's query (and merge where the row has it) accuracy, workload and sketch_config:
   • 74 rows are identical.
   • 8 seeded rows (kll-percall, kll-cdf, kll-dynamic, hydra-kll, hydra-kll-cdf, coco, nitro-count-min, nitro-count-sketch, all lib) give different numbers on two runs of main itself when no seed is set. With seed=7 they match main exactly.
   • 9 rows also give different numbers on two runs of main itself: hydra-cms/lib, hydra-cs/lib, hydra-univmon-entropy/lib and /polars, univmon-cardinality/lib, univmon-entropy/lib and /oxide, univmon-two-tier-cardinality/lib, univmon-two-tier-entropy/lib. The branch values fall within main's own run-to-run spread.
2. Prefix property: aqpbm-cli/tests/trials.rs (a_later_start_reproduces_the_same_trials).
3. relabel keeps stream_items, stream_distinct and stream_l2, and Count-Min aae_top1 differs across trials (relabel_keeps_the_stream_and_moves_the_collisions).
4. reseed: KLL rank_err_* differ across trials and are identical on a rerun. Count-Min under reseed is rejected with the list of seeded rows.
5. A fresh trial matches a --trials 1 --spec <its descriptor> run for CMS (i64) and Bloom (string).
6. hll lg_k=14, N = 2^16, 1024 trials: 276.8 s in one process vs 262.1 s as 1024 single-trial processes. Both give identical records. One process is not faster. About 200 ms of each trial is the query measurement itself: the cardinality comparator asks the sketch 4096 times. Process start costs about 20 ms. Before this change, a process that did not set BENCH_WARMUP_SECS=0 also paid the 10 s default warmup.
7. cargo fmt --check, cargo clippy --workspace --all-targets --all-features -- -D warnings and cargo test --workspace pass.

Different from the spec:

• UnivMonQParams has seed_index, not seed, so univmon-q rows are not reseed-able.
• --trials with --workload-spec is rejected.
• A key column that already carries key_offset is rejected under fresh/relabel.
• reseed is checked per (variant, library): the oxide KLL rows (no seed) and the polars baselines (seed ignored) are rejected.
• Every trial option counts as a trial run, --trials 1 --trial-start S included, so the accuracy-only checks apply to it.

🤖 Generated with Claude Code