Parse a key-value log line

Constraints

1. 1A field is a key of letters, digits and underscores, an equals sign, and a value. The key never starts with a digit.
2. 2A value wrapped in double quotes runs to the closing quote and may contain spaces. An unquoted value ends at the next space.
3. 3Anything in the line that is not a pair — a timestamp, a level, a bracketed component, a trailing comment — is ignored.
4. 4Coerce an unquoted value: true and false become booleans, a whole number becomes an int, a decimal becomes a float, and everything else stays a string.
5. 5An unquoted value that is empty or a single dash means nothing was set, and becomes None.
6. 6A quoted value is always text. An empty quoted string is an empty string, not None, and a quoted number stays a string.
7. 7When a key appears more than once, the last value wins.
8. 8A line that is not a string returns an empty dict rather than raising.

Worked example

Take the first case line: level=INFO status=200 latency=13.5 msg="job failed on shard 3" ok=true retries=0 The message value contains three spaces. Splitting the line on whitespace produces msg="job as one token, and the rest of the message is scattered across tokens with no equals sign in them — so the field that says what went wrong is the one field that is lost. retries=0 is the other trap. Coercing it and then testing the result for truthiness treats zero retries as no value, and the field disappears. Zero is a measurement; only an empty value and a dash mean nothing was set.

What this tests

Reading a format by its grammar rather than by splitting on a convenient character, and the difference between a value that is absent and a value that is falsy. Both mistakes produce output that looks fine until the one line that matters.