marque-engine

Pipeline orchestration for Marque — wires the scanner, parser, and rule sets into a configured, ready-to-run engine.

marque-engine is the middle of the pipeline. It owns the Engine (single document) and BatchEngine (concurrent, async) APIs that the CLI, the WASM build, and the server all sit on top of.

Role in Marque

marque-config ─┐
marque-core ───┼─→ marque-engine ─→ LintResult / FixResult + audit log
marque-rules ──┘

The engine drives the scanner, parses each candidate, evaluates every registered rule with a RuleContext, and — for fix mode — promotes FixProposals to AppliedFix records when their confidence meets the configured threshold.

Usage

Single-document lint:

use marque_capco::capco_rules;
use marque_config::Config;
use marque_engine::Engine;

let engine = Engine::new(Config::default(), vec![Box::new(capco_rules())]);
let result = engine.lint(b"(S) example text");

println!("errors: {}, warnings: {}", result.error_count(), result.warn_count());

Apply fixes with audit records:

use marque_engine::{Engine, FixMode};
# use marque_capco::capco_rules;
# use marque_config::Config;
# let engine = Engine::new(Config::default(), vec![Box::new(capco_rules())]);

let result = engine.fix_with_threshold(b"(S) example", FixMode::Apply, None)?;
for applied in &result.applied {
    // serialize `applied` to the NDJSON audit stream
}
# Ok::<(), marque_engine::InvalidThreshold>(())

Behavior

• Confidence gate. Only FixProposals whose confidence is at or above the configured threshold are promoted. Proposals below the gate are surfaced as suggestions in LintResult, never written.
• Reverse-span application. Fixes are sorted and applied from the end of the document to the start, so earlier spans remain valid as edits land. Overlapping fixes are rejected per the C-1 invariant.
• Page context. The engine builds a marque_ism::PageContext from preceding portion markings and hands an Arc<PageContext> to banner/CAB rules through RuleContext. The accumulator resets at scanner-emitted MarkingType::PageBreak candidates (form-feed \f and \n\n\n+), before attempting to parse the page-break candidate, so a malformed candidate cannot block the reset.
• Audit log. In production code paths, AppliedFix is constructed exclusively by the engine via the __engine_promote constructor. Rule crates must not bypass this. Test code MAY call __engine_promote directly to construct synthetic fixtures for audit-emission unit tests, scoped per Constitution V Principle V (#[cfg(test)] only, never commingled with engine output, and test-fixture construction only — not a general-purpose convenience). See the architectural-invariants section of the workspace CLAUDE.md and the __engine_promote doc comment.
• Clock injection. Engine::with_clock accepts a Clock impl; a FixedClock is provided for deterministic test snapshots and for the CLI --fixed-timestamp debug flag.

Batch processing

With the batch feature, BatchEngine wraps Engine behind Arc and uses recoco-utils semaphores for row + byte backpressure. CPU-bound work runs on tokio::task::spawn_blocking. Results stream out in completion order, not submission order — correlate by the echoed id.

Features

A cache feature (LMDB-backed incremental cache, opt-in) is planned for v0.2 but is not yet present in Cargo.toml.

WASM compatibility

The sync Engine path is WASM-safe; build with default-features = false to drop the async/batch dependencies. marque-wasm consumes the engine this way.

License

Marque License 1.0 (LicenseRef-MarqueLicense-1.0). See LICENSE.md.