Module identifiers 

Normative identifier extraction for compaction retention (INV-C2).

Anchored-iterative compaction replaces the conversation’s anchor summary wholesale, so an identifier the summary drops is dropped forever. The retention invariant (INV-C2, conformance row CONF-2) is stated over extractable identifiers — and this module is the single, normative definition of that term. The compaction recall eval (#1134) and the mint-time retention gate (#1135) both call this extractor; neither may ship a private variant, or the eval would measure a different property than the gate enforces.

Token classes

An extractable identifier is a token (or token run) in one of five classes, matching the conformance suite’s ambiguity resolution:

• IdentifierClass::Url — an http:// or https:// token, trailing punctuation trimmed.
• IdentifierClass::Amount — a numeral with a currency symbol ($/€/£), a trailing %, a following unit word from UNIT_WORDS, or at least four digits (order numbers, PINs).
• IdentifierClass::OpaqueId — a machine-shaped token: at least five characters drawn from [A-Za-z0-9._/-] containing both a letter and a digit (UUIDs, ord_93k2f7x, ZK-4471-BQ).
• IdentifierClass::ProperNoun — a run of two or more consecutive capitalized words, leading English function words stripped (Mirela Okafor, not The Mirela).
• IdentifierClass::Quoted — the content of a straight double-quoted span of 3–120 bytes containing a letter. This is how an open question stays trackable: prose that tags it (open item: "night berthing at dock 7") makes the tag extractable, while free interrogative prose is semantic content the recall eval measures but the gate cannot extract.

Survival predicate

An identifier survives into a candidate text when the candidate contains its text verbatim (case-sensitive substring — is_retained). A paraphrase (Okafor, Mirela) does not count: the gate’s failure mode on over-strictness is rejecting a candidate summary and keeping the prior anchor, which is the safe direction (INV-C3).

Structs

Identifier

One extracted identifier: its class and its verbatim text.

Enums

IdentifierClass

The class of an extracted identifier. See the module docs for the normative grammar of each class.

Constants

UNIT_WORDS

Unit words that promote a bare numeral to an IdentifierClass::Amount.

Functions

extract_identifiers

Extract every identifier in text, deduplicated and in a stable (class, text) order.

is_retained

The normative survival predicate: candidate retains identifier when it contains the identifier’s text verbatim (case-sensitive substring).

missing_identifiers

The identifiers extracted from prior that do NOT survive (per is_retained) into candidate.