Add "bytes read" metric to per-prompt statistics

[laanwj]

Add a "bytes read" metric to per-prompt statistics. The "/stats" command in chat already shows this, but the stats for "waste run" didn't.

[mfethe1]

Independent CPU/synthetic-container audit at 656af11ed49a83fce74792a3d58c139a57ae53b9 (base c66c7b3febf261590abea540e74602d8e7759d62) found two small but useful follow-ups:

1. The stats format string contains GB read\]. Clang emits unknown escape sequence '\]'; removing that backslash removes the new warning. Base had no warnings under the same default flags.
2. The new sum is callback/decode-only, whereas the existing engine stats include work outside those callbacks. Chunked prefill precedes the per-token bb baseline, so summing callback deltas cannot count those reads. Using the same label for these different scopes makes the measurement easy to misinterpret.

On a tiny synthetic container, callback bytes were 0 versus engine bytes 282624; adding the omitted term in the measurement control closed the difference, and a fixed-seed repeat reproduced it. That is not a general omission percentage: this fixture fits entirely in cache, so it cannot establish the magnitude under decode-time cache pressure. Whole-run accounting also needs explicit before/after boundaries rather than assuming an absolute counter is per-request.

Project make check stayed 60 passed / 0 failed / 14 skipped on both base and PR.

Would you prefer the deliberately narrower label decode GB read, or a whole-request metric with a before/after counter boundary and regression coverage? I can contribute a focused companion patch once the intended metric is settled, rather than create a competing implementation.

[mfethe1]

Concrete minimal option for the scope question: keep the existing callback sum, make its narrow scope explicit, and correct the unit label (the divisor is 1ULL << 30, so GiB, not decimal GB). This avoids changing accounting semantics while you decide whether a whole-request metric belongs in a separate change.

Applied locally on this PR's unchanged head 656af11ed49a83fce74792a3d58c139a57ae53b9:

diff --git a/cli/main.c b/cli/main.c
index b19d2e7..ae916f6 100644
--- a/cli/main.c
+++ b/cli/main.c
@@ -876,7 +876,7 @@ static int run_segs(waste_ctx *c, const opts *o, const seg *segs, int ns,
     if (show_stats && s.n) {
         const double sec = s.ms / 1000.0;
         fprintf(stderr, "\n[%u tokens, %.2f s, %.2f tok/s | experts %llu hit / "
-                        "%llu miss = %.0f%% | %.2f GB read\]\n",
+                        "%llu miss = %.0f%% | %.2f decode GiB read]\n",
                 s.n, sec, s.n / sec,
                 (unsigned long long)s.hit, (unsigned long long)s.miss,
                 100.0 * (double)s.hit / (double)(s.hit + s.miss ? s.hit + s.miss : 1),

Verification on macOS CPU: fresh default make -j2 succeeds with zero compiler warnings; make check gives 55 passed / 0 failed / 18 skipped in this environment. The skip set includes missing uv and external tokenizer/model fixtures; this is not the previous environment's 60-pass run and not real-checkpoint validation. The edit changes only the printed label and removes the invalid C escape; it does not change the accumulated counter, fix omitted prefill reads, or claim whole-request accounting.

No competing branch/PR was pushed. This is an independently built, tested suggestion for the author to incorporate if callback/decode scope is intended. AI-assisted contribution.