bench: restore the rows a removed cache ceiling was still excluding (#100)

[Xof]

Closes #100 (3 findings, DESIGN/docs).

Stacked on #134 → #133 → #132 → #131 → #130 → #129 → main.

BENCH-5 — coverage lost to a premise that no longer holds

micro_grid skipped cells and dropped two whole rows:

• TX_BUDGET_BYTES = 8 MiB, justified as avoiding "Chisel's CacheFull at large
  1000-per-tx writes (cache hard ceiling is ~16 MB)"
• update-1000pertx and delete-1000pertx never registered, because "1000
  random updates/deletes pin ~1000 distinct dirty data pages, exceeding
  Chisel's 2048-page cache ceiling. The cells are not measurable under default
  cache settings."

Both halves are stale. CACHE_SIZE_PAGES is 256, not 2048, and the strict
cache ceiling went away when the bench engines gained the spillway at
production-default scale — runner.rs already says "there is no strict
cache-page ceiling on transaction size" and chisel_engine.rs sets
spillway_max_bytes(cache_max_bytes * 1024).

Measured rather than assumed

Every excluded cell, against the current engine at CACHE_SIZE_PAGES:

cell	per-tx bytes	result
allocate-1000pertx 16KB	16 MB	0.71 s
allocate-1000pertx 128KB	128 MB	2.70 s
allocate-1000pertx 1MB	1 GB	18.51 s
update-1000pertx 16KB	—	1.27 s
delete-1000pertx 16KB	—	0.80 s

All complete. Capacity is no longer the constraint — wall clock is, for exactly
one cell.

What I did, and where I deviated

TX_BUDGET_BYTES is kept but reframed as a bench-runtime budget and raised
to 128 MiB. That admits the 16KB and 128KB columns (real coverage lost for no
reason) and still excludes 1MB × 1000: at ~18.5 s per iteration across five
engine modes, that single cell would add hours to every micro-grid run.

The issue's direction says to delete TX_BUDGET_BYTES outright. That is the
one part I did not follow, and the comment now records the measurement behind
the decision so it can be revisited on evidence rather than re-derived. If you
want the 1MB × 1000 column too, say so and I'll drop the constant — it is a
one-line change.

update-1000pertx and delete-1000pertx are registered.

BENCH-6 — the documented grid size was off by 105 cells

The header claimed a "270-cell micro grid ... 9 row groups"; six rows
registered 165 cells. A reader checking that a run completed would conclude 105
cells had silently failed — and the live-but-uncalled seed_for arms for the
two missing rows reinforced the illusion that they ran.

The header now states 230 cells across 8 rows with the per-row breakdown,
and runner.rs's "8 of 9 rows: rows 1, 2, 4–9" is replaced by what actually
distinguishes them (every row except read-warm, which reuses one engine and
has its own capture function).

Verified two independent ways:

1. Derived from the registration rules — and the same derivation reproduces the
   old 165 against the old constants, which is what validates the model.
2. Counted directly: cargo bench --bench micro_grid -- --list → 230.

The seed_for arms are now live again rather than deleted.

BENCH-7 — the published methodology footer understated Chisel 3×

cross-engine.md's footer told readers "Each engine takes a single fsync per
commit through the disk write cache." Chisel performs three (pre-drain,
data pages, superblock) — as src/lib.rs, lifecycle.rs and the bench's own
runner.rs all state.

This renderer's module doc describes its output as suitable for the README and
1.0 release notes, so the wrong claim propagates into public performance
material, and it invites the reader to assume the three commit paths are
equivalent when Chisel is doing strictly more syscall work for the same
throughput number. The footer now states each engine's protocol: Chisel's three
fsyncs, redb's Durability::Immediate, and SQLite's WAL + synchronous=FULL +
fullfsync=ON.

Note on runtime

Re-registering the two rows and widening the budget takes the grid from 165 to
230 cells, so a full cargo bench --bench micro_grid will take meaningfully
longer. That is the coverage the issue is about restoring, but it is a real
cost worth being aware of. bench/results/ is untracked, so no committed
artifact goes stale.

Verification

cargo test all green, cargo clippy --workspace --all-targets -- -D warnings
clean, cargo fmt --check clean.