db_url: "" currently silently derives the state_dir-based sqlite URL,
which is a default applied to a SET value; state_dir: "" already
aborts. Failing test first, fix follows.
An explicitly-null key (port: null, bare port:, debug: ~, and every
other config key including metrics subkeys) is a SET value under the
no-silent-fallback rule and must abort startup naming the key, instead
of silently taking the default as it does today. All 13 subtests fail
against the current behavior; the fix follows.
A config value that is set but unparseable or invalid now aborts
startup with an error naming the offending key and value; defaults
apply only to omitted keys. Unknown top-level config keys and unknown
metrics subkeys abort startup naming each unknown key, so typos like
whitelist_hosts fail immediately instead of being silently ignored. A
config file that exists at a standard location but fails to parse is
now a fatal error instead of being skipped with a warning. state_dir
is verified creatable and writable with a probe file before the
listener binds. Port must be in 1-65535 (fractional values are
rejected, not truncated), upstream_connections_per_host must be at
least 1, allowlist_hosts entries must be bare hostnames, sentry_dsn
must be a valid URL when set, and metrics credentials must be set
together. The stale signing_key comment in config.example.yml (keyless
mode was never implemented) now states the actual requirement.
TODO.md records the completed step per its Workflow section.
Encode the required fail-fast behavior as tests ahead of the
implementation: a config value that is SET but unparseable or invalid
must abort startup (defaults apply only to OMITTED keys), unknown
top-level keys and unknown metrics subkeys must abort naming the key,
a malformed config file at a standard location must abort instead of
being skipped with a warning, and state_dir must be creatable and
writable at startup. Mechanically extracts newFromSmartConfig from
config.New so the construction path is testable without fx; current
lenient behavior is unchanged, so the new enforcement tests fail.
