Module duration_parse 

The single duration grammar for <number><unit> strings.

Every subsystem that accepts a human-written duration (“5m”, “200ms”, “1h”) parses it here, under one set of rules:

• A unit suffix is required. "30" is an error, not an implicit millisecond count — a unitless number reads as seconds to most people and as milliseconds to most APIs, so guessing silently misreads it.
• The vocabulary is ms, s, m, h, d, w, matched case-insensitively, tolerating whitespace before the suffix.
• Arithmetic is checked. An oversized value is reported, never clamped: a timeout that silently becomes u64::MAX presents to the user as a hang, which is far harder to diagnose than a parse error.

These rules were previously three forked copies that disagreed. Most sharply, when_budget.timeout was parsed by both the CLI manifest validator and the runtime trigger_register, which disagreed on overflow — the same string was rejected at validation but accepted and clamped at registration. One grammar removes that class of divergence.

parse_millis owns the arithmetic and returns a structured DurationParseError; each caller maps that onto its own error type and wording, which stays the caller’s business.

The float/long-form cache-TTL parser (llm::cache) and the OptionsParser millis path (stdlib::options, which rejects unit strings outright) are deliberate outliers and do not use this module.

Enums

DurationParseError

Why a duration string could not be interpreted.

Functions

parse_millis

Parse a <number><unit> duration string into milliseconds.