feat: alignment pairs a misread sentence with its other reading; a suspect names its witness

spec/05 §5.6 gains a substitution step: an unpaired sentence beside one at
least 0.8 alike (2L/(|a|+|b|) over the lossy projections' code points) is
paired with it, so a misread is one number, equation or word suspect rather
than a dropped and an added one. §5.7 and suspect.schema.json gain an
optional `witness`, checked by doc.suspect.witness. alignment.jsonl gains a
`substitutions` case; the first case is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

CHANGELOG.md

@@ -35,3 +35,14 @@ Packaging, commands and lazy imports only; the formats stay at 0.1.0, and no com
 - `src/` is Apache-2.0, as `tools/` was, and the package metadata states `Apache-2.0 AND CC0-1.0` for the code and the data it carries.
 - The pixi workspace installs the package editable, and a `dev` environment builds the sdist and the wheel with `pixi run -e dev dist`.
 - Pushing a version tag publishes the package to the Gitea package registry through `.gitea/workflows/publish.yml` ([deploy/](deploy/README.md#publish-the-python-package)); a release tag versions the package, and [CONTRIBUTING.md](CONTRIBUTING.md#releasing) says how that relates to the format version.
+
+### Unreleased
+
+Two changes asked for by the first producer that compares a converted reading with a second witness (paperstract with a JATS file).
+Both are minor changes in the sense of [spec/00](spec/00-introduction.md): no valid file becomes invalid, and no hash, ID or key changes; the next release numbers them.
+
+- Alignment has a substitution step ([spec/05 §5.6](spec/05-hashes.md#56-alignment), step 3): in each stretch between pairs, an unpaired sentence of the first reading is paired with the first unpaired sentence of the second whose similarity (2*L* / (|*a*| + |*b*|) over code points of the lossy projections, *L* their longest common subsequence) is at least 0.8, both of prose length 20 or more.
+  A sentence one reading misread is now one `number`, `equation` or `word` suspect rather than a `dropped` and an `added` one, so a producer re-aligning the same two readings can get different suspects than before; the pairs and suspects of readings with no such sentence are unchanged, as the first alignment vector is.
+  `paperspec.align` gains `lcs_length` (bit-parallel, Hyyrö 2004) and `similarity`, and `alignment.jsonl` a second case, `substitutions`: a misread number, a misread word, a sentence too short to be paired this way and two sentences too unlike to be one.
+- A suspect may name the witness its second reading came from (`witness`, a witness ID; [spec/05 §5.7](spec/05-hashes.md#57-suspects), `suspect.schema.json`), and a package whose suspect names a witness it does not have breaks `doc.suspect.witness`.
+  No broken package exercises the rule yet: adding one changes the count of broken packages that the 0.1.0 entry above states, so it waits for the release.