chore: drop bioconda recipe drafts; add 2.0.1 release notes

The bioconda recipe drafts under bioconda/ were a staging area for the bioconda
PRs; both are now merged upstream (salmon 2.0 via bioconda-recipes #66280, the
final-C++ salmon/1.12 subrecipe via #66337), so the in-repo copies are redundant
and would only drift from the canonical recipes. The salmon-cpp draft is dead
(that approach was dropped for the subrecipe).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: rob-p

docs/release-notes-2.0.1.md

@@ -0,0 +1,53 @@
+# salmon 2.0.1
+
+A patch release with two internal accuracy improvements to selective-alignment
+quantification. **No breaking changes, no new options, and no output-format
+changes** — indices built with 2.0.0 work unchanged (no rebuild needed), and
+`quant.sf` / inferential-replicate outputs are produced exactly as before.
+
+## What changed
+
+Both changes refine how fragments are weighted across **multimapping**
+transcripts (paralogs and alternative isoforms), bringing the estimates closer to
+C++ salmon's:
+
+- **Fragment-length probability in the equivalence-class weights.** The
+  selective-alignment equivalence-class weights now include the fragment-length
+  term (salmon's `logFragProb`), which had been omitted. Without it the
+  conditional weights were flattened for transcripts whose implied insert size
+  differs, slightly degrading how shared reads are split.
+- **Abundance-aware fragment-length-distribution training.** The fragment-length
+  distribution is now trained from fragments sampled by their abundance-aware
+  posterior (matching salmon), so reads shared between near-duplicate transcripts
+  contribute the dominant transcript's implied length. This concentrates the
+  learned distribution as C++ salmon does.
+
+## Impact
+
+On benchmark data the per-transcript agreement with C++ salmon improves, and the
+remaining difference sits at the run-to-run noise floor:
+
+- Real data (GEUVADIS `ERR188044`, GRCh38 cDNA): per-transcript Spearman vs C++
+  salmon 0.9705 → **0.9710**; NumReads Pearson **0.9996**; mapping rate
+  unchanged.
+- Simulated ground truth (human, polyester): the per-transcript Spearman gap to
+  C++ salmon shrinks to ≤ 0.002 across difficulty levels and both the default
+  (VBEM) and `--useEM` optimizers — at/near the noise floor.
+
+Performance and memory are unchanged, and runs remain reproducible (run-to-run
+variation is unaffected). Existing 2.0.0 results are still valid; re-quantifying
+picks up the slightly improved estimates.
+
+## Installation
+
+Unchanged from 2.0.0:
+
+```sh
+curl --proto '=https' --tlsv1.2 -LsSf \
+  https://github.com/COMBINE-lab/salmon/releases/latest/download/salmon-cli-installer.sh | sh
+# or: cargo install salmon-cli   |   conda install -c bioconda -c conda-forge salmon
+```
+
+## Documentation
+
+Full docs at **<https://combine-lab.github.io/salmon>**.