fix(bench): fail loud on a missing latency percentile instead of defaulting to 0

[ELares]

Why

The last c7g #676 run was wasted by an "empty p999" -- and the harness reported a result anyway.

Root cause is a silent default that can't fire:

p999="$(json_field "${open_json}" p999_us)"
[[ -n "${p999}" ]] || p999="0"

jq -r .missing_key prints the string "null", which is non-empty, so the -n guard never triggers. A missing percentile became "null" (or 0 via the sed fallback) and flowed into the report and the moat ratio as a plausible-looking 0us tail.

That's the worst failure mode: the number looks real, so a paid bench run is silently invalidated.

What

Adds json_num FILE KEY LABEL -- requires a non-negative integer, else aborts naming the file, key, offending value, and dumping the first 20 lines of the results file. The four open-loop percentiles now use it: the p99.9 is the #518 moat metric and the #676 A/B verdict, so it must never be defaulted.

Verified

Unit-checked standalone against all five shapes: good file -> 30000 (rc=0); missing key / JSON null / non-numeric / empty file -> each exits 1 with a FATAL diagnostic. bash -n clean.

Pre-flight for the sanctioned c7g run, so a bad extraction aborts in seconds instead of producing a garbage tail.

[github-actions]

perf-gate (A5)

Same-runner ratchet of HEAD against the merge-base (both rebuilt and measured in this job).
PASS = within the noise band, WARN = a real move inside budget (does not fail), FAIL = past budget in the bad direction.

metric	base	head	delta%	band	budget	verdict
qps_median (peak)	82455.98	82336.02	-0.15%	+/-5.00%	drop <= 15%	PASS
bytes_per_key int	47.59	46.98	-1.28%	det	rise <= 5%	PASS
bytes_per_key embstr	60.86	60.83	-0.05%	det	rise <= 5%	PASS
bytes_per_key raw	332.61	332.78	0.05%	det	rise <= 5%	PASS

Overall: PASS

• qps: noisy on shared CI, so the band comes from the base reps spread (floored at 5%); a drop is only a regression past the 15% budget.
• bytes_per_key: deterministic (allocator-true memmodel), so a tight 5% rise budget; any rise beyond it FAILs.
• Open-loop tails / criterion micro-benches are reported-not-failed (tail noise is high) and are not part of this ratchet.
• An intentional perf trade is landed by raising the relevant budget in this PR with a documented reason (CI never auto-commits a baseline).