bug: --full flag silently ignored during benchmark — per-transaction .csv.zst never written

[russfellows]

Bug: --full flag silently ignored — per-transaction .csv.zst file never written

Description

Running any benchmark with the --full flag does not produce a per-transaction
.csv.zst output file. Only the .json.zst aggregate is written, making post-hoc
per-operation analysis impossible. The flag appears to be accepted without error but
has no effect.

Steps to Reproduce

warp put --host=<host> --access-key=<key> --secret-key=<secret> --full

Expected: Two output files written after the benchmark:

• warp-put-<timestamp>.csv.zst — one row per operation (full per-transaction log)
• warp-put-<timestamp>.json.zst — aggregated summary

Actual: Only warp-put-<timestamp>.json.zst is written. The .csv.zst file is
never created.

Root Cause

Introduced in commit 483f844, which added addCollector(). That function was
written to gate per-transaction collection on ctx.Bool("full"), but the --full
flag was only registered in analyzeFlags — never added to benchFlags. As a
result, ctx.Bool("full") always returns false during benchmark execution, and
the collector always falls back to EmptyOpsCollector (no ops stored, no file
written).

// addCollector() — intended gate:
if ctx.Bool("full") {        // always false during benchmark — flag not registered
    // fan ops to NewOpsCollector + LiveCollector
} else {
    return bench.EmptyOpsCollector, updates  // always taken
}

Impact

• warp analyze --full <file> (full per-operation latency/throughput analysis) is
  effectively unreachable — there is no .csv.zst file to analyze.
• Users explicitly requesting --full for detailed post-benchmark analysis get
  silently degraded output with no warning or error.
• warp merge, warp cmp, and per-request filtering (--analyze.op,
  --analyze.host, --analyze.skip) all require a .csv.zst — none of these
  workflows are usable.

Additional Finding

Running warp analyze file.csv.zst without --full re-aggregates individual
operations into 1-second buckets before reporting — producing results equivalent to
reading the .json.zst aggregate — with no indication to the user that per-operation
precision has been lost.

Fix

A fix is available in PR #475, which:

1. Makes addCollector() correctly gate on --full (flag already present in
   analyzeFlags, which is combined with benchFlags for all benchmark commands).
2. Makes --full additive: both .csv.zst and .json.zst are always written
   together when the flag is set.
3. Mirrors the fix in benchserver.go for distributed mode.
4. Adds a warning in warp analyze when a .csv.zst is analyzed without --full,
   directing users to the more accurate analysis path.
5. Updates README documentation and adds 7 unit tests.

PR: #475