Add tiny lfu to roadmap (#17)

* Update policy documentation to reflect planned and implemented policies

- Revised the list of upcoming cache policies in the README to include GDSF and TinyLFU/W-TinyLFU.
- Added TinyLFU/W-TinyLFU to the implemented policy summaries in the roadmap documentation, ensuring clarity on current and future features.

* Update policy documentation to correct upcoming policies list

- Removed outdated references to ARC and CAR from the planned policies section in the README.
- Ensured the list accurately reflects the current focus on LIRS, GDSF, and TinyLFU/W-TinyLFU.

Author: TKorr

docs/policies/README.md

@@ -58,7 +58,7 @@ Enable the corresponding feature flag for each policy. See [Compatibility and Fe
 
 ### Roadmap Policies (Planned)
 
-See [Policy roadmap](roadmap/README.md) for upcoming policies (ARC, CAR, LIRS, etc.).
+See [Policy roadmap](roadmap/README.md) for upcoming policies (LIRS, GDSF, TinyLFU/W-TinyLFU, etc.).
 
 ### Implemented Policy Summaries (Short)
 

docs/policies/roadmap/README.md

@@ -14,4 +14,5 @@ If a policy moves into production code, its document should be moved back to
 - [LFU Aging](lfu-aging.md)
 - [LIRS](lirs.md)
 - [OPT](opt.md)
+- [TinyLFU / W-TinyLFU](tinylfu.md)
 - [TTL](ttl.md)

docs/policies/roadmap/tinylfu.md

@@ -0,0 +1,45 @@
+# TinyLFU / W-TinyLFU (Admission + Eviction Design)
+
+## Goal
+Improve hit rate under skewed and one-hit-wonder workloads by using
+frequency-based admission instead of admitting every miss.
+
+## Core Idea
+TinyLFU is primarily an **admission policy**:
+- Keep an approximate recent frequency sketch (usually Count-Min Sketch).
+- On a miss, compare candidate frequency against a sampled victim.
+- Admit only if the candidate appears "hotter" than the victim.
+
+W-TinyLFU combines:
+- **Window cache** (small recency-focused region, typically LRU)
+- **Main cache** (segmented/protected region)
+- **TinyLFU admission gate** between window and main
+
+This keeps recent bursts responsive while avoiding long-tail pollution.
+
+## Core Data Structures (Typical)
+- Hash index `K -> Entry`
+- Window segment (e.g., LRU list or ring)
+- Main segment (e.g., SLRU-style probation/protected)
+- Frequency sketch (Count-Min Sketch)
+- Optional reset/aging counter for sketch decay
+
+## Complexity & Overhead
+- Access/update in sketch: O(1) with small constant factors
+- Admission decision: O(1)
+- Extra memory for sketch and segmented metadata
+- Approximate counts can produce false positives, but usually good tradeoff
+
+## Notes For CacheKit
+- Fits best as **policy composition**: storage + segmented eviction + admission.
+- Keep hot-path updates allocation-free (pre-sized sketch, fixed segments).
+- Make admission optional/configurable for apples-to-apples benchmarks.
+- Benchmark against `S3-FIFO`, `ARC/CAR`, and `Heap-LFU` on:
+  - Zipfian
+  - scan + point lookup mixes
+  - shifting hotspots
+
+## References
+- Einziger et al. (2017): "TinyLFU: A Highly Efficient Cache Admission Policy".
+- Caffeine design notes (W-TinyLFU implementation details).
+- Wikipedia: https://en.wikipedia.org/wiki/Cache_replacement_policies