Diagnostics and limits

A query can handle a missing label with a default. A division by zero still fails. Zirql distinguishes absent data, unknown domain facts and errors so that handling one does not hide another.

Outcome	Example	What the query can do
Absence	lookup({}, "word")	Use a default or match presence
Unknown predicate	The host cannot resolve a source reference	Retain the host’s uncertainty rules
Error	1 div 0, incompatible types or exhausted work	Fail with a location

default handles absence. It does not catch arithmetic or resource errors. Negating an unknown predicate keeps it unknown. A filter retains only known true answers; a required native match guard or path target refuses an unknown final answer.

Query and source locations

Query diagnostics point into the original expression. A missing field points to its access, a failed arithmetic operation to that operation, and a limit failure to the work that exhausted the allowance. Source failures retain the host’s source coordinates and document identity.

A request can fail before it executes. Preflight checks reject undefined names, known incompatible types, duplicate bindings and effects in pure contexts. A condition that is always false does not hide those errors. Missing data or a division by zero in an unselected runtime branch is not evaluated.

When a function fails, the host can report a bounded context of active calls. Tail-call reuse records the current activation rather than every completed recursive call. Diagnostic wording and presentation belong to the host.

Request allowances

The embedding application chooses finite parsing and execution limits:

Allowance	What it bounds
Work	Cumulative evaluation, validation, scans and other charged operations
Results	New or copied values and container slots under a program budget
Stream items	The size of one native stream
Nesting	Logical expression nesting, with a library maximum of 64

A nested query shares the remaining allowance. A new input stream, function call or comprehension does not start a fresh budget. These are accounting limits, not promises about elapsed time or peak memory in bytes.

What uses the allowance

Comprehensions visit every candidate, including candidates rejected by a filter. Reducers receive eager arrays. Sorting inspects keys and uses temporary storage; grouping builds records and member arrays. Copying a saved result or rendering it can spend additional work and result units.

A short expression can therefore do substantial work. For example, two independent generators over words visit every pair, while a generator over positions with an access to the next position visits only adjacent pairs. Choose the expression that matches the question you intend to ask.

Spent work and results are not refunded when a value is released or an operation fails. Active scope and nesting state are restored after a failure. An exhausted request fails; it does not quietly return a shortened report or partial path search.

Function calls and depth

The application also sets an active function-depth limit. A compatible tail call can reuse an activation; a call with pending work needs another one. Each call still spends work. An endless tail recursion eventually fails its work allowance. Expression nesting and native callback depth remain separate limits.

Effects and output

Parsing and validation are read-only. Pure expressions and functions cannot edit or emit. A host with explicit output or editing operations defines their effect and rollback policy. A buffered command may publish only a complete result; a live emitter may already have produced output before a later failure.

Adapter authors should read budget and diagnostic APIs for error categories, charging rules, callback obligations and cleanup. The execution guide explains preparation, sharing and measured costs.