fix: startup-time numbers printed as literal SLF4J specifiers (1.1.1)

[ddsha441981]

Correctness pass found by running the published 1.1.0 artifact unmodified
against spring-petclinic (Spring Boot 4.1.0, Framework 7.0.8, Java 25.0.2,
470 beans). No API, schema, or configuration changes.

The bug

Three log templates used {:+d} / {:.1f}. SLF4J interpolates only {}, so
the specifier was printed verbatim and every following argument shifted by one
slot:

⏱ Startup Time: 6687ms -> ms ({:+d}ms, {:.1f}% 7178)   # before
⏱ Startup Time: 6687ms -> 7418ms (+731ms, 10.9% slower) # after

It hit the diff summary, its no-regression variant, and the startup-time
gate's log.error — the last being the line a CI reader lands on first. Numbers
are now pre-formatted through signed(long) / oneDecimal(double) and passed as
plain {} arguments. The WireDoctorRegressionException message was already
correct and is unchanged.

Also: startupTimePercentChange in wiredoctor-gate.status was written with a
default-locale String.format("%.1f", …), so a comma-decimal build machine
emitted 4,3 into a file that is meant to be machine-read. It now goes through
oneDecimal(…), which pins Locale.ROOT.

The docs bug, which is the more damaging one

ci-gating.md told you to record the baseline with ./mvnw spring-boot:run
while Step 3 gates on java -jar. devtools is on the classpath for the former
and excluded from the repackaged jar for the latter — on spring-petclinic that
single difference is 12 removed beans and a 31% startup-time delta, enough to
trip the startup-time gate on a build where nobody changed a line of code. The
quickstart prescribed the exact mismatch that produces a false failure.

Step 1 now records from the jar, with a callout on keeping profiles and
spring.main.web-application-type identical across both runs. Step 3 gains a
warning that a Maven-goal gate cannot fail a build at all: with devtools
present, spring-boot:run runs main on its own restart thread, so the gate
trips, BUILD SUCCESS prints, and the job exits 0.

configuration.md documents output-path for projects whose build lints the
source tree (nohttp rejects the http:// license headers vendored in the
self-contained report, failing the next build), and that scan-packages also
keeps WireDoctor's own beans out of the smell rankings. _config.yml declares
the warning callout the docs reference.

Verification

End-to-end against the petclinic jar with a locally built WireDoctor — both the
summary line and the armed gate:

WireDoctor regression gate 'startup-time' tripped: startup time increased by
290ms (4.3%) vs baseline jar-baseline.json (6687ms -> 6977ms).
Thresholds: >=100ms AND >=2%

Gate run exits 1; wiredoctor-gate.status carries startupTimePercentChange=4.3.

257 tests green — 246 autoconfigure + 11 actuator, +4 new. Among them a guard
that scans every log.* template in the module for a non-{} specifier, so the
whole defect class stays closed. Verified it actually fails by re-injecting
{:+d} before restoring.

Note on publishing

1.1.1 is a source tag only — nothing is published to Maven Central by this
PR, so install snippets in README.md, docs/index.md and
docs/why-wiredoctor.md deliberately stay on 1.1.0. The valuable half of this
change is the docs, and that ships via GitHub Pages the moment this merges.