mirror of
https://github.com/shankar0123/certctl.git
synced 2026-06-07 23:31:39 +00:00
c4ed3da30b4fa6e420540ceee12d33a9fac548eb
7 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 shankar0123
|
1279172e9b |
loadtest: close Phase 8 SCALE-H2 — add scale-tier scenarios
Phase 8 of the certctl architecture diligence remediation closes
SCALE-H2 by adding three new k6 scenarios that exercise the scale-
relevant load surfaces the API tier + connector tier left uncovered:
fleet-scale bulk renewal, ACME enrollment burst, and agent heartbeat
storm.
Audit miscount + path correction (live-grep at Phase 8 audit time)
==================================================================
- The Phase 8 prompt referenced both `deploy/test/load/` and
`deploy/test/loadtest/`. Repo truth: the existing harness lives at
`deploy/test/loadtest/`. New scenarios land there.
- The audit's prior framing "k6 covers the API tier at 50 req/s
only" omitted Bundle 10 (2026-05-02) which added four connector-
tier handshake scenarios (nginx/apache/haproxy/f5) at 100 conns/min
each, plus the Phase 5 ACME directory/nonce/ARI scenario at 100 VUs
in `k6/acme_flow.js`. Phase 8 appends to what's there rather than
rewriting.
What ships
==========
Three new k6 scenario files under deploy/test/loadtest/k6/:
bulk_renewal.js — 10K-cert seed + 5 req/s POST /bulk-renew × 5min
p99 < 5s, p95 < 2s, errors < 1%
acme_burst.js — 200 VU sustained × directory/nonce/ARI × 5min
directory p95 < 500ms, nonce p95 < 300ms,
renewal-info p95 < 800ms, 5xx-only < 0.1%
Pins RFC 7807 rate-limit response shape via
acme_rate_limit_shape_ok Counter.
agent_storm.js — 5K-agent seed + 167 req/s POST /heartbeat × 5min
p99 < 1s, p95 < 500ms, errors < 0.1%
Two seed SQL fixtures under deploy/test/loadtest/seed/:
01_bulk_renewal_certs.sql — 10,000 managed_certificates rows
linked to seed_demo.sql FKs (iss-local, o-alice, t-platform,
rp-standard). status='active', expires_at distributed across
next 30 days, name prefix `loadtest-bulk-` so the scenario
can scope its criteria. Idempotent via
ON CONFLICT (name) DO NOTHING.
02_agent_fleet.sql — 5,000 agents rows with name prefix
`loadtest-agent-`. status='Online', last_heartbeat_at
staggered across prior 60s, OS distribution 80%/10%/10%
linux/windows/darwin. Idempotent via
ON CONFLICT (id) DO NOTHING.
Plus seed/README.md documenting the opt-in profile + when these
run vs the default `make loadtest` fast path.
Compose + Makefile + CI wiring
==============================
deploy/test/loadtest/docker-compose.yml gains four new services,
all gated behind the `scale` compose profile so the default
`make loadtest` is unchanged:
scale-seed — one-shot postgres:16-alpine container that runs
every ./seed/*.sql in lexical order against the
same postgres the server uses. Depends on
postgres healthy + certctl-server healthy (so
migrations + seed_demo.sql have already run).
k6-scale-bulk — grafana/k6:0.54.0 driver running bulk_renewal.js
k6-scale-acme — grafana/k6:0.54.0 driver running acme_burst.js
k6-scale-agent — grafana/k6:0.54.0 driver running agent_storm.js
Each driver depends_on scale-seed completed_successfully so the
scenarios never run against an unseeded DB (the acme scenario
doesn't need the seed itself but uses the same dependency chain for
ordering predictability).
Makefile gains four new phony targets:
loadtest-scale-bulk - runs bulk_renewal.js via compose --profile scale
loadtest-scale-acme - runs acme_burst.js
loadtest-scale-agent - runs agent_storm.js
loadtest-scale - all three serially
.github/workflows/loadtest.yml gains a new k6-scale matrix job that
runs after the existing k6 job (needs: k6) with a matrix on the
three scenarios — fail-fast: false so a regression in one scenario
doesn't cancel the others. Same workflow_dispatch + weekly cron
cadence as the existing API + connector tier job.
Documentation
=============
docs/operator/scale.md gains a new "Scale-tier scenarios (SCALE-H2,
Phase 8)" section between the cursor-pagination subsection and the
profiling-production subsection. Documents:
- Scenario + seed + sustained load table
- Threshold contract (regression guards, NOT measured baselines)
- Measured-baseline table with TBD placeholders + the canonical-
hardware capture procedure
- How to run the scale tier locally
- Four documented limitations (JWS-signed ACME, scheduler renewal
scan throughput, production-sized Postgres, pull-only deployment
model)
deploy/test/loadtest/README.md gains a short "Scale tier (Phase 8
SCALE-H2, 2026-05-14)" section pointing at scale.md as the canonical
operator-facing baseline source. Avoids duplication; the README
remains the harness-mechanics doc.
Deliberate deviations from the prompt
======================================
The Phase 8 prompt's "concrete deliverables" section referenced
`deploy/test/load/` (no -test) for the new k6 files. The actual
harness lives at `deploy/test/loadtest/` — the new files land there
to match existing convention. The prompt's audit-questions section
also referenced `deploy/test/loadtest/` so the prompt was internally
inconsistent on this; repo truth wins.
The prompt described the ACME burst as "200 concurrent ACME orders
against /acme/profile/<id>/new-order ... pin the rate-limit response
shape." new-order is JWS-signed (RFC 8555 §7.4 requires JWS for
every POST except newAccount-pre-account-key flows). k6 doesn't
ship JWS and bundling a signer (e.g. lego) into the k6 container
would obscure the server-side latency the scenario is trying to
measure. Same trade-off the existing Phase 5 acme_flow.js made.
Phase 8's acme_burst.js measures the unauthenticated
directory + nonce + ARI surface at burst rate AND pins the 429
rate-limit response shape via a custom Counter that increments only
when the response is `application/problem+json` with the
`urn:ietf:params:acme:error:rateLimited` type. End-to-end JWS
conformance under load remains a follow-up; the canonical JWS
correctness gate is `make acme-rfc-conformance-test` (lego-based,
non-load).
Deferred (operator-side, not engineering)
==========================================
Canonical-hardware baseline capture. The TBD placeholders in
docs/operator/scale.md's measured-baseline table are intentional —
sandbox-captured numbers from a developer laptop are misleading
(same anti-pattern the original loadtest README guards against).
Operator triggers loadtest.yml from the Actions tab, waits for the
k6-scale matrix jobs to complete, downloads the per-scenario
summary artifacts, copies p50/p95/p99 into the table, commits the
captured numbers alongside the date + commit SHA.
Files changed (10):
.github/workflows/loadtest.yml (+72 -1)
Makefile (+47 -1)
deploy/test/loadtest/README.md (+28 -1)
deploy/test/loadtest/docker-compose.yml (+108 -1)
deploy/test/loadtest/k6/bulk_renewal.js (new, 106 lines)
deploy/test/loadtest/k6/acme_burst.js (new, 192 lines)
deploy/test/loadtest/k6/agent_storm.js (new, 124 lines)
deploy/test/loadtest/seed/01_bulk_renewal_certs.sql (new, 95 lines)
deploy/test/loadtest/seed/02_agent_fleet.sql (new, 92 lines)
deploy/test/loadtest/seed/README.md (new, 86 lines)
docs/operator/scale.md (+109 -0)
Verification (sandbox-runnable):
python3 -c 'import yaml; yaml.safe_load(open("deploy/test/loadtest/docker-compose.yml"))'
→ compose YAML OK
python3 -c 'import yaml; yaml.safe_load(open(".github/workflows/loadtest.yml"))'
→ workflow YAML OK
grep -E 'bulk_renewal|acme_burst|agent_storm' deploy/test/loadtest/k6/*.js
→ all three scenarios + tags present
grep loadtest-scale Makefile
→ 4 new targets registered in .PHONY + 3 recipes + 1 aggregate
Runtime verification (deferred — requires docker on canonical hardware):
make loadtest-scale-bulk # 10K cert fixture + 5 req/s × 5min
make loadtest-scale-acme # 200 VU × 5min
make loadtest-scale-agent # 5K agent fixture + 167 req/s × 5min
make loadtest-scale # all three serially
Closes: cowork/certctl-architecture-diligence-audit.html#fix-SCALE-H2
|
||
|
shankar0123
|
75097909e9 | |||
|
shankar0123
|
0f81c1b956 |
ci: re-fix CodeQL #32 + repair loadtest f5-mock build context
Two unrelated CI failures from run #25305811340; fixed in one
commit since neither needs the other to land first.
CodeQL alert #32 (go/log-injection at middleware.go:68) reopened
after
|
||
|
shankar0123
|
bee47f0318 |
acme-server: cert-manager integration test + production hardening (Phase 5/7)
Closes the production-readiness loop on the ACME surface. After this
commit, certctl ships per-account rate limits + a GC sweeper for
expired ACME state + a kind-driven cert-manager 1.15 integration test
+ a lego-driven RFC conformance harness + a k6 loadtest scenario for
the unauthenticated ACME path.
Architecture:
- Rate limits live in-memory + per-replica. Restart wipes the
counters; orders/hour caps are eventual-consistency anyway. A
3-replica certctl-server fleet behind an LB effectively has 3x
the configured throughput per account; persistent rate limiting
is a follow-up if production telemetry shows abuse patterns we
can't catch in a single restart cycle. Per-key + per-action
isolation: ActionNewOrder/acc-1, ActionKeyChange/acc-1, and
ActionChallengeRespond/<challenge-id> are independent buckets.
- GC loop follows the existing scheduler-loop pattern (atomic.Bool
+ sync.WaitGroup; see crlGenerationLoop for shape). Three
independent SQL sweeps per tick (DELETE expired nonces; UPDATE
pending authzs whose expires_at < now() to expired; UPDATE
pending/ready/processing orders whose expires_at < now() to
invalid). Each sweep is a single statement; failures are logged-
and-continued so a failing nonces sweep doesn't block authzs.
Per-sweep 1m timeout bounds a stuck Postgres.
- cert-manager integration test is gated on KIND_AVAILABLE so CI
skips it cleanly (kind is too heavy for per-PR). Operators run
locally via 'make acme-cert-manager-test'; the harness brings up
a fresh cluster each run + tears it down on Cleanup.
- lego conformance harness drives a real ACME client through
register → run → cert-PEM-landed against a hermetic certctl
stack. Catches RFC-shape regressions third-party clients would
hit before they ship.
- k6 ACME-flow scenario hammers the unauthenticated surface
(directory + new-nonce + ARI synthetic-id) at 100 VUs × 5m. JWS-
signed flows are out of scope for k6 (no JWS support); they're
covered by the lego harness above.
What ships:
- internal/api/acme/ratelimit.go (+ ratelimit_test.go: 7 cases —
disable-when-perHour-zero, capacity, per-key isolation, per-
action isolation, refill-over-time, RetryAfter, concurrent-access
with -race + 200 goroutines × 200 calls).
- internal/repository/postgres/acme.go: 4 new methods —
CountActiveOrdersByAccount + GCExpiredNonces + GCExpireAuthorizations
+ GCInvalidateExpiredOrders. Each a single SQL statement.
- internal/service/acme.go: SetRateLimiter + GarbageCollect +
rate-limit gates at 3 entry points (CreateOrder + RotateAccountKey
+ RespondToChallenge) + concurrent-orders gate at CreateOrder.
2 new sentinels (ErrACMERateLimited, ErrACMEConcurrentOrdersExceeded);
5 new GC metrics (gc_runs / gc_run_failures / gc_nonces_reaped /
gc_authzs_expired / gc_orders_invalidated).
- internal/scheduler/scheduler.go: ACMEGarbageCollector interface +
acmeGCRunning atomic.Bool + acmeGCInterval + 2 setters (SetACME-
GarbageCollector + SetACMEGCInterval) + acmeGCLoop following the
crlGenerationLoop shape.
- internal/api/handler/acme.go: writeServiceError gains rateLimited
(429 + RFC 8555 §6.7) + concurrent-orders-exceeded mappings.
- internal/config/config.go: 5 new env vars
(CERTCTL_ACME_SERVER_RATE_LIMIT_ORDERS_PER_HOUR=100,
CERTCTL_ACME_SERVER_RATE_LIMIT_CONCURRENT_ORDERS=5,
CERTCTL_ACME_SERVER_RATE_LIMIT_KEY_CHANGE_PER_HOUR=5,
CERTCTL_ACME_SERVER_RATE_LIMIT_CHALLENGE_RESPONDS_PER_HOUR=60,
CERTCTL_ACME_SERVER_GC_INTERVAL=1m).
- cmd/server/main.go: NewRateLimiter() + SetRateLimiter() at
startup; conditional SetACMEGarbageCollector(acmeService) +
SetACMEGCInterval(cfg.ACMEServer.GCInterval) when Enabled+
GCInterval > 0.
- deploy/test/acme-integration/: kind-config.yaml + cert-manager-
install.sh + clusterissuer-trust-authenticated.yaml +
clusterissuer-challenge.yaml + certificate-test.yaml + conformance-
lego.sh + certmanager_test.go (//go:build integration + KIND_AVAILABLE
gate).
- deploy/test/loadtest/k6/acme_flow.js + README ACME-flows section.
- Makefile: 2 new PHONY targets (acme-cert-manager-test +
acme-rfc-conformance-test).
- docs/acme-server.md: status flipped to Phase 5; Configuration
table grows 5 rows; new 'Phase 5 — operational guidance' section
explaining rate-limit math + GC sweeper semantics + cert-manager
integration + lego conformance + k6 baseline.
Tests:
- 'go vet ./...' clean across the repo.
- 'go test -short -count=1 ./internal/...' green across every
affected package (service / acme / handler / scheduler / repo /
config).
- 'go vet -tags=integration ./deploy/test/acme-integration/' clean
(the integration test compiles cleanly with the build tag).
- The kind/cert-manager harness is gated behind KIND_AVAILABLE so
CI skips by default; operators run locally via 'make acme-cert-
manager-test'.
Engineering history: cowork/WORKSPACE-CHANGELOG.md 'ACME-Server-5'.
|
||
|
shankar0123
|
c26cef37a1 |
loadtest: capture sandbox-aggregate placeholder for API-tier baseline
Closes Top-10 fix #2 of the 2026-05-02 deployment-target audit re-run (see cowork/deployment-target-audit-2026-05-02-rerun/RESULTS.md). Replaces the four TBD cells in deploy/test/loadtest/README.md ## Current baseline with a sandbox-aggregate placeholder so the README isn't lying about having a baseline section ready to diff against. Numbers (both rows show the same aggregate — see footnote): p50=2.12 ms, p95=6.19 ms, p99=8.58 ms, error rate 0.00% (1002 requests, 100.15 req/s sustained, 0 failures across 10s) Capture environment, called out explicitly in the new methodology block: - Linux/aarch64 unprivileged sandbox (NOT canonical hardware) - Postgres 14.22 native (NOT 16-alpine in compose) - 10s scenarios (NOT 5 minutes) - Both rows have the same numbers because the sandbox run did not emit per-scenario tagged metrics in summary.json — the threshold contract still expects per-scenario p95/p99 from a canonical run. Footnote ([^1]) frames these as a sanity floor, not the per-scenario baseline the threshold contract is written against. The follow-up canonical capture via `gh workflow run loadtest.yml` on the GitHub-hosted ubuntu-latest runner will replace these with real per-scenario numbers (and will keep the canonical methodology block that's already pinned below). Connector-tier table (## Connector-tier captured baseline) is intentionally left at TBD: that block explicitly anti-patterns committing numbers without a Docker-equipped canonical run, and the sandbox can't run the four target sidecars. No code changes; doc-only. Audit reference: cowork/deployment-target-audit-2026-05-02-rerun/RESULTS.md Top-10 fix #2. |
||
|
shankar0123
|
e292faafc6 |
loadtest: per-connector deploy throughput scenarios + target sidecars + README baseline section
Closes Bundle 10 of the 2026-05-02 deployment-target coverage audit
(see cowork/deployment-target-audit-2026-05-02/RESULTS.md). Pre-fix,
deploy/test/loadtest/k6.js drove only the API-tier throughput path
(POST /api/v1/certificates + GET /api/v1/certificates) — the operator-
facing rate at which an automation client can submit cert requests.
The deploy hot path (cert deployed to a target — connector-tier
latency) had no benchmarks. Procurement asks "can certctl handle our
5,000-NGINX fleet at 47-day rotation?" and the answer should be a
number with methodology, not a claim.
This commit ships v1 of the connector-tier loadtest harness:
1. Target-side sidecars added to docker-compose.yml: nginx-target,
apache-target, haproxy-target, f5-mock-target. Each daemon serves
a starter cert (ECDSA P-256, multi-SAN) written into a shared
./fixtures/target-certs/ volume by a new target-tls-init
container. f5-mock-target re-uses the in-tree
deploy/test/f5-mock-icontrol/ image (already used by the deploy-
vendor-e2e CI job) and generates its own self-signed cert via
tls.go::selfSignedCert at startup.
2. Fixture configs committed under deploy/test/loadtest/fixtures/:
- nginx.conf — minimal HTTPS server, single 200 OK location.
- httpd.conf — self-contained Apache config with the minimum
module set + SSL vhost.
- haproxy.cfg — minimal SSL-terminating frontend backed by a
static "ok" backend.
3. k6 scenarios added (4 new): nginx_handshake, apache_handshake,
haproxy_handshake, f5_handshake. Each runs constant-arrival-rate
at 100 conns/min for 5 minutes. Latency captured by k6's
http_req_duration metric covers TCP connect + TLS handshake +
tiny HTTP request/response — that's the end-to-end "connection
readiness" latency a deploy connector cares about.
4. summary.json gains a connector_tier object with per-target
p50/p95/p99/max/avg/error_rate/iterations breakdowns. Operators
tracking a connector regression diff connector_tier.<type>
between runs. Implementation: a new enrichWithConnectorTier
helper that reads data.metrics keyed by target_type tag and
shallow-merges the breakdown into the summary before
serialisation.
5. Threshold contract per target type:
- nginx/apache/haproxy: p99 < 3s, p95 < 1s.
- f5-mock: p99 < 5s, p95 < 1.5s (iControl REST
handler does slightly more work per
request than pure TLS termination).
- All scenarios: error rate < 1% (k6 default; any 4xx/5xx
counts as failed).
Any change pushing past these fails the workflow.
6. README documents the methodology + the baseline-number table for
the connector tier. Numeric values are em-dash placeholders
pending the first clean canonical-hardware run; the accompanying
commit message in that follow-up captures the methodology line
alongside the numbers. Out-of-scope is documented explicitly:
- Full agent-driven deploy poll loop (POST cert with target
binding → poll deployments endpoint → verify served cert).
v2 of the harness — needs the agent registration + target-
binding API surface plumbed end-to-end in the loadtest stack.
- Kubernetes target via kind-in-docker. kind requires
`privileged: true` and is operationally fragile in CI;
deferred until Bundle 2 (real k8s.io/client-go) lands and a
CI-friendly envtest harness is wired.
- Real F5 BIG-IP. CI uses the in-tree f5-mock; real-appliance
benchmarking is out of scope.
7. CI workflow .github/workflows/loadtest.yml timeout-minutes
bumped from 15 to 25. The harness now boots four additional
target sidecars before the k6 run; their healthchecks add
~30-60s. The k6 scenarios themselves are still 5 minutes (run
in parallel, not serially). 25 minutes absorbs that plus slow
CI runners and cold image caches without letting a stuck
container consume the runner indefinitely. Trigger remains
workflow_dispatch + cron — sustained 25-minute runs are too
slow for per-PR signal.
What this connector tier explicitly does NOT measure (documented in
the k6.js header + README):
- The agent-driven full deploy hot path (v2 follow-up).
- K8s target (Bundle 2 dependency).
- Real F5 appliance.
- Issuer-side throughput (handled by issuer-coverage-audit fix #8).
Verified locally:
- python3 -c "import yaml; yaml.safe_load(...)" on docker-compose.yml
and .github/workflows/loadtest.yml — clean.
- node -c on k6.js — clean syntax.
- gofmt / go vet on the rest of the tree (no Go diff in this commit).
- Manual smoke against docker-compose pending — operator validates
on the canonical-hardware first run; if any fixture config is off,
fix-up commit lands separately so the methodology change and the
numeric baseline have independent reviewability.
No Go code changes; this is a loadtest-harness-only commit.
Audit reference: cowork/deployment-target-audit-2026-05-02/RESULTS.md
Bundle 10.
|
||
|
shankar0123
|
3a665ae6ba |
loadtest: add k6 harness for certctl API throughput
Closes the #8 acquisition-readiness blocker from the 2026-05-01 issuer coverage audit. Pre-fix, certctl had zero benchmarks or load tests for any API path. An acquirer evaluating "can certctl handle our 50k-cert fleet at 47-day rotation" had nothing to point at; CA/B Forum SC-081v3 lands 47-day TLS in 2029, and operators need real numbers, not hand- waved capacity claims. What landed: - deploy/test/loadtest/docker-compose.yml — minimal stack (postgres + tls-init bootstrap + certctl-server with CERTCTL_DEMO_SEED=true so the FK rows the script needs exist + grafana/k6:0.54.0 driver). Pinned k6 version so threshold expressions stay stable across runs. k6 command runs the script once and exits with the threshold-driven exit code so `--exit-code-from k6` propagates non-zero on any regression. - deploy/test/loadtest/k6.js — two scenarios at 50 req/s × 5 min, staggered 5s. Scenario 1: POST /api/v1/certificates (issuance- acceptance hot path: auth + JSON decode + validation + service CreateCertificate + DB insert). Scenario 2: GET /api/v1/certificates (most-trafficked read endpoint, exercises pagination). Hard thresholds: p99 < 5s + p95 < 2s for issuance-acceptance, p99 < 2s + p95 < 800ms for list, error rate < 1% globally. constant-arrival- rate executor (NOT constant-vus) so VU-bound load doesn't backpressure the offered rate and mask capacity ceilings. __ENV.CERTCTL_BASE lets the same script run on the operator's workstation (https://localhost:8443) and inside the compose stack (https://certctl-server:8443). - deploy/test/loadtest/README.md — documents what's measured (API tier: auth → DB) vs what's NOT (issuer connector latency: pinned separately by certctl_issuance_duration_seconds from audit fix #4; full ACME enrollment flow: deferred — sustained 100/s through multi-RTT pebble takes pebble tuning + crypto helpers k6 doesn't ship with). Threshold contract pinned. Baseline numbers row reads TBD until the operator captures on a representative workstation; methodology pinned so future tuning commits land alongside refreshed baselines that are diffable. - deploy/test/loadtest/.gitignore — results/{summary.json,summary.txt} + certs/ (per-run TLS bootstrap output). Both regenerate on every run; committing them would create huge per-run diffs. - deploy/test/loadtest/results/.gitkeep — placeholder so the directory exists in fresh checkouts (the k6 container mounts it). - Makefile: new `loadtest` target spinning up the compose stack with --abort-on-container-exit --exit-code-from k6 and printing the summary. Added to .PHONY + help. Explicitly NOT in `make verify` — load tests are minutes long and don't gate per-PR signal. - .github/workflows/loadtest.yml — workflow_dispatch (manual) + weekly cron at Mon 06:00 UTC. NOT per-push. 15-minute hard cap. Always uploads results/ as an artifact (90d retention) so a regression has a diffable artifact even when k6 exited non-zero. Read-only repo permissions. - docs/architecture.md: new "Performance Characteristics" section citing the harness location, scenarios, thresholds, scope (what's measured vs not), and where the captured baseline lives. Inserted before the existing "What's Next" section. Scope decisions documented in the README + this commit message: - The audit prompt's k6 example targeted POST /api/v1/certificates + ACME-via-pebble. CreateCertificate exercises auth + DB but the downstream issuer-connector call is async (renewal scheduler); that's the right surface for "request-acceptance" throughput. Driving the connectors directly would load-test someone else's API. - Pebble was excluded from the harness stack. Sustained 100/s through ACME's order/challenge/finalize flow needs pebble tuning + k6 crypto helpers that don't exist out of the box. README flags this as a deferred follow-up. Acquirer impact: the diligence question "what's your throughput?" now has a number with a reproducible methodology and a regression guard, not a claim. The first operator run captures the baseline into README.md so subsequent tuning commits are diffable. Verified locally: - gofmt -l . clean - go vet ./... clean - staticcheck ./... clean - go build ./... clean - bash scripts/ci-guards/H-1-encryption-key-min-length.sh — clean (the 38-byte loadtest key is above the 32-byte floor) - bash scripts/ci-guards/openapi-handler-parity.sh — clean - bash scripts/ci-guards/test-compose-scep-coherence.sh — clean - make -n loadtest produces the expected command sequence - The first `make loadtest` run from the operator's workstation populates the README baseline numbers (committed in a follow-up). Audit reference: cowork/issuer-coverage-audit-2026-05-01/RESULTS.md Top-10 fix #8. |