KS78: Recency tie-breaker for score equality (#13)

[Liorrr]

Summary

• Adds microsecond-precision recency epsilon to break score ties in favor of newer memories
• recency_epsilon = created_at.timestamp_micros() * 1e-18 — negligible impact on normal scoring
• Applied after all boosts/caps but before final sort
• Fixed test to properly distinguish Google-only vs Meta memory matches

Changes

• crates/shrimpk-memory/src/echo.rs: +14/-2 lines (recency epsilon + test fix only)

Test plan

• cargo test -p shrimpk-memory — all non-ignored tests pass
• cargo clippy -p shrimpk-memory -- -D warnings — clean
• cargo check --workspace — clean

Closes #13

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR adds a microsecond-precision recency epsilon (step 7c7) to the EchoEngine scoring pipeline to break score ties in favor of newer memories (KS78 / issue #13). It also fixes a test bug where the "Google-only" memory was being incorrectly matched because the Meta memory's content also contains the word "Google".

Key changes:

• A negligible epsilon timestamp_micros * 1e-18 is added to each result's final_score after all boosts and the inflation cap, so that among equal-scoring memories the newer one surfaces first.
• The test recency_boost_newer_memory_ranks_higher is fixed: the predicate for the "Google-only" result now correctly excludes entries that also contain "Meta".
• The epsilon magnitude for current timestamps is ~0.00174 (absolute), but the difference between any two entries created within the same year is at most ~3 × 10⁻⁵ — genuinely negligible relative to any meaningful similarity gap.

Confidence Score: 5/5

Safe to merge — no panics, no memory-safety issues, epsilon math is correct, and the test fix is accurate.

The epsilon formula is mathematically sound: the i64→f64 cast is lossless at current timestamp magnitudes (~1.744e15 μs, ULP ≈ 0.25 μs), and the maximum epsilon difference between any two memories created within the same year is ~3×10⁻⁵ — genuinely negligible relative to meaningful similarity gaps. The only concern is a minor documentation gap at the cap site (P2), not a behavioral bug.

No files require special attention; the single P2 is a comment-only suggestion at crates/shrimpk-memory/src/echo.rs line 1620.

Important Files Changed

Filename	Overview
crates/shrimpk-memory/src/echo.rs	Adds 9-line recency epsilon loop (step 7c7) after the inflation cap, plus a one-line test predicate fix; math is sound, no panics, minor cap-bypass worth noting.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["7c2: Temporal query boost"] --> B["7c3: Topic-label boost"]
    B --> C["7c4: Preference-update multiplier"]
    C --> D["7c5: Career/intro adjustment"]
    D --> E["7c6: Inflation cap\n(final_score ≤ similarity + 0.50)"]
    E --> F["7c7 NEW: Recency epsilon\n(+= timestamp_micros × 1e-18)"]
    F --> G["7d: Re-sort by final_score DESC"]
    G --> H["7d2: Subject-diversity cap"]
    H --> I["7e: Optional reranker"]

Loading

_{Reviews (1): Last reviewed commit: "fix: revert accidental hebbian_boosts ch..." | Re-trigger Greptile}

[greptile-apps]

Liorrr has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.

[Liorrr]

Closing: branch contaminated with config changes from other worktrees. Recreating as clean PR.