docs: CONTRACT.md reflects the unbiased cap + features path; add CHANGELOG

CONTRACT.md: the four-verbs scope now names both variants --index and
features --index (same engine, same envelope); the depth-cap description is
updated from "deterministic downsampling" to the unbiased content-hash reservoir
(no silent variant drops; over_max_depth surfaced); features added to the
"what's bounded" list. New CHANGELOG.md documenting the v0.1.0 capabilities.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: logannye

CHANGELOG.md

@@ -0,0 +1,46 @@
+# Changelog
+
+All notable changes to Rosalind are recorded here. Versions follow [Semantic Versioning](https://semver.org).
+
+## [0.1.0] — 2026-06-02
+
+First tagged release: a deterministic, low-memory genomics engine where **memory is a verifiable
+contract** — predict it before you commit, honor it during the run, and verify it after.
+
+### The memory contract
+- `rosalind plan` — predict a job's peak memory against a declared budget *before* committing a byte.
+- `rosalind variants … --enforce` — honor the budget: refuse up front (exit 3) or fail loud (exit 4),
+  never a silent OOM-kill. Record-only without `--enforce`.
+- `rosalind verify` — re-check a run's BLAKE3 receipt without re-running.
+- The **`rosalind-budget` GitHub Action** — enforce the contract in *your* CI (fail the build on breach).
+
+### Variant calling
+- **Bounded whole-genome germline SNV calling** (`variants --index`) over a coordinate-sorted BAM and a
+  persisted index: peak memory tracks coverage, not BAM size. Calibrated, abstention-aware.
+- **Unbiased depth-cap downsampling** — a content-hash reservoir that bounds the working set without
+  biasing allele balance (no silent variant drops); dropped-read counts surfaced in the receipt.
+- **Tumor/normal somatic** SNV + simple-indel calling (`somatic`).
+- **Measured detection accuracy** on simulated diploid truth (precision/recall 1.00/1.00 on clean data);
+  `eval-germline` / `eval-somatic` truth-set comparison (the `eval-germline` path is GIAB-ready).
+
+### Index, alignment, I/O
+- Build-once, memory-mapped, byte-reproducible FM-index (`index` / `locate`).
+- Single-contig FM-index aligner (`align`); deterministic external-merge coordinate sort (`sort`);
+  streaming gzip/bgzf FASTA/FASTQ input.
+
+### Reproducible ML feature substrate
+- `rosalind features --index` — the same bounded stream as a per-locus feature **TSV**, byte-identical
+  run-to-run, with a hash receipt: **bit-reproducible training inputs**.
+- `python/rosalind.py` (stdlib + numpy) loads it; `examples/reproducible_features_demo.py` proves the
+  bit-reproducibility end-to-end.
+
+### Reproducibility & distribution
+- Byte-identical primary outputs and a canonical-JSON BLAKE3 receipt per run.
+- Prebuilt static/native binaries via a tagged release + `install.sh`; a 60-second quickstart.
+
+### Research direction
+- The `~√t` (square-root-space) framework is retained as honest framing for **sublinear-space index
+  construction** (Phase D); today's index build is `O(reference)` and the contract covers call/query.
+  See [`docs/OPEN_PROBLEMS.md`](docs/OPEN_PROBLEMS.md).
+
+[0.1.0]: https://github.com/logannye/rosalind/releases/tag/v0.1.0

CONTRACT.md

@@ -14,7 +14,7 @@ the Phase-D research direction, not a present claim.)
 
 ## The four verbs
 
-The contract applies to the bounded whole-genome germline path, `rosalind variants --index`.
+The contract applies to the bounded whole-genome paths — `rosalind variants --index` (germline calling) and `rosalind features --index` (per-locus feature egress), which share the same streaming engine and therefore the same `plan`/`--enforce`/`verify` envelope.
 
 ### 1. Declare
 
@@ -48,8 +48,10 @@ With `--enforce`:
 - otherwise the run completes within budget.
 
 Without `--enforce`, the budget is **record-only**: the run always completes and the verdict is recorded in
-the receipt. The active read set is capped at `--max-depth` (default 1000; `0` = uncapped) — deterministic
-downsampling that bounds the working set; output changes only at sites deeper than the cap.
+the receipt. The active read set is capped at `--max-depth` (default 1000; `0` = uncapped) by an
+**unbiased** content-hash reservoir — it bounds the working set without biasing allele balance, so a deep
+variant is *not* silently dropped. Output changes only at sites deeper than the cap, and the dropped-read
+count is surfaced (stderr + the receipt's `over_max_depth`).
 
 ### 4. Verify — `rosalind verify`
 
@@ -66,8 +68,9 @@ auditability story containers can't give you for a non-deterministic caller.
 
 ## What's bounded (honest scope)
 
-- **Germline `variants --index`** is the bounded path: peak ≈ the largest contig's reference + the
-  depth-capped active set, **independent of BAM size**. Reads stream one record at a time.
+- **Germline `variants --index`** and **`features --index`** are the bounded paths: peak ≈ the largest
+  contig's reference + the depth-capped active set, **independent of BAM size**. Reads stream one record at
+  a time; output rows (VCF calls or feature rows) stream straight to disk with no genome-wide buffer.
 - **Somatic** (`somatic`) is **region-bounded**, not whole-genome-bounded (it collects both pileup streams
   for the region).
 - **Index *build*** (`rosalind index`) is **O(reference)** in RAM today; `plan --reference` reports an