Report errors, memory and cost alongside the latency numbers

[lgoyal6]

The C02 columns from .agent-work/wt-c02c08. bench/ was untouched by #19, so this applied without conflict.

Why

A latency table never says three things on its own: what failed while the numbers were being produced, how much memory the process held, and what the work cost. A p999 next to an unreported delivery failure is not a measurement of the transport, it is a measurement of the messages that happened to arrive.

Errors, per role, counted rather than dropped:

• The receiver silently returned on a payload too short to hold the header, so a framing bug would have surfaced as a slightly lower delivery ratio and nothing else.
• A sequence that does not advance is a duplicate or a reorder. The receiver cannot tell which from one number, so it reports the count and does not claim to know.
• Offered minus received is the error a latency table hides best: a message that never arrived contributes no sample, so dropping it improves every percentile printed above it.
• The request-reply client is closed loop with one outstanding, so the reply to request N must be N. A stale or duplicate reply still set got_reply and still pushed a sample, timed from the wrong request. That mismatch is counted, and so is the wait cap firing, which the old loop passed over silently: a wedged transport produced a short run and a clean-looking percentile table.
• The sender's error is offered work it never handed to the transport before the wall cap; reporting only sent hides the shortfall.

Memory is peak and steady RSS together. ru_maxrss is bytes on Darwin and kilobytes on Linux, and getting that wrong is silently wrong by three orders of magnitude in the flattering direction on exactly one of the two platforms, so the trap is named in a comment instead of guessed at.

Cost is CPU seconds, the axis that matters for a poll/tick transport with no event-loop driver: both roles spin, so wall time says how long the run lasted and CPU time says what it burned to last that long. Involuntary context switches sit next to it so a tail from a busy machine can be told apart from a tail from the transport. The USD figure is arithmetic on cpu_s, so the rate is printed with it.

resource.h is deliberately separate from common.h: common.h is the wire format and schedule shared by the taut, TCP and ENet binaries and documented in docs/BENCHMARKS.md, while this is process accounting and writes its own CSV beside the latency one, so no documented schema changes.

Verification, and its limits

bench/ is Linux-only in CMake and no CI job builds it, so neither the normal local build nor CI compiles this file. Rather than land it unchecked:

• latency_bench.cc compiles clean under the project's full strict set (-Werror -Wpedantic -Wshadow -Wconversion -Wsign-conversion -Wold-style-cast -Wcast-align -Wdouble-promotion) on this machine.
• Both platform branches of resource.h were compiled and run, not just the local one. The Darwin branch reads a real RSS and emits a well-formed CSV row; the Linux branch was compiled by forcing its arm of the conditional and exercised for header, row and path behaviour. resource_path rewrites the extension as documented.
• The library suite is unaffected: 61/61.

What is not verified is an end-to-end benchmark run over real sockets, which needs Linux and netem.

🤖 Generated with Claude Code