Struct FidelityScorer 

pub struct FidelityScorer;

Stateless heuristic scorer that assigns and applies fidelity levels to a message window.

Call FidelityScorer::score_and_apply after apply_prepared_context() returns to enforce the three-level representation (Full / Compressed / Placeholder) on historical messages. The scorer never touches exempt messages (INV-07 through INV-10).

Examples

use zeph_context::fidelity::{FidelityConfig, FidelityScorer};

let scorer = FidelityScorer;
let cfg = FidelityConfig { enabled: false, ..FidelityConfig::default() };
// With `enabled = false` the scorer is a no-op.
let mut messages = vec![];
scorer.score_and_apply(&mut messages, "query", &[], &cfg, &MockTc, 0, false, None, None).await;

struct MockTc;
impl zeph_common::memory::TokenCounting for MockTc {
    fn count_tokens(&self, text: &str) -> usize { text.len() / 4 }
    fn count_tool_schema_tokens(&self, _: &serde_json::Value) -> usize { 0 }
}

Implementations

impl FidelityScorer

pub async fn score_and_apply( &self, messages: &mut Vec<Message>, query: &str, planned_tools: &[PlannedToolHint], config: &FidelityConfig, tc: &dyn TokenCounting, inserted_count: usize, allow_upgrade: bool, embed_provider: Option<&dyn LlmProviderDyn>, compress_provider: Option<&dyn LlmProviderDyn>, )

Score all non-exempt messages and apply fidelity rendering in-place.

Steps (per spec §5 data flow):

1. Guard: return early when enabled == false.
2. Build exempt set (INV-07 through INV-10).
3. Score each non-exempt message with normalized weight sum (INV-05).
4. Apply floor invariant: a message with a persisted fidelity tag cannot be upgraded to a less restrictive level unless allow_upgrade = true.
5. Apply tool-pair atomicity — both get min(score_a, score_b) (INV-03).
6. Render Compressed / Placeholder messages (INV-12).
7. Merge consecutive same-role Placeholder messages (INV-04).

Parameters

• messages — mutable message window (includes system prompt at index 0).
• query — current user query; drives semantic signal.
• planned_tools — DAG lookahead hints; empty slice disables plan signal.
• config — scoring thresholds and weights.
• tc — token counter used for Placeholder/Compressed rendering.
• inserted_count — number of memory messages freshly injected at indices 1..1+inserted_count; these are always exempt (INV-10).
• allow_upgrade — when true the persisted floor invariant is bypassed. Pass true only from the proactive regrade path; pass false everywhere else.
• embed_provider — optional LLM provider used for query and per-message embeddings when config.semantic_scoring_provider is set. Pass None to fall back to keyword overlap.
• compress_provider — optional LLM provider used for Compressed rendering when config.compress_provider is set. Pass None to fall back to truncation.

The two providers may be the same instance or different ones (e.g. a fast embedding model for embed_provider and a generative model for compress_provider).