Remediate Ε-W3-T1 package-compat-fixtures review findings

Freeze *.rsbank as binary via .gitattributes; add a truncated additive_forward fixture proving the exact-size proof beats TooNew; enumerate the fixture dir to catch orphaned files; make fixture-size checks fatal instead of just logged; pin fixture version asserts as literals, not build-relative.

tests/fixtures/package_compat/README.md

@@ -12,20 +12,29 @@ The reason is the whole point of the corpus. These bytes exist to catch a format
 that quietly breaks a compatibility direction. A fixture regenerated by the build that
 broke it agrees with that build by construction and catches nothing — which is exactly
 the failure mode a version ladder exists to prevent. The same argument forbids a test
-that builds its own fixture at run time.
+that builds its own fixture at run time. The repo-root `.gitattributes` (`*.rsbank
+binary`) keeps this mechanical: without it, git's NUL-sniffing heuristic could
+text-classify a future short/ASCII fixture and CRLF-mangle a line ending on a Windows
+checkout, silently breaking the frozen-bytes premise.
 
-If a fixture stops decoding, the answer is never to re-cut the fixture. Either the format
-change was structural (bump `minReaderVersion`, add a new fixture, and leave the old one
-asserting the refusal) or it is a regression.
+A fixture's BYTES are frozen forever; a fixture's ASSERTION is not. `additive_forward.rsbank`
+and `refuse_structural.rsbank` carry version pairs one step past THIS build's ladder (2/1
+and 2/2). When a future build's own `kPackageFormatVersion` reaches 2, `refuse_structural.rsbank`
+classifies `Readable` under the new ladder — its bytes never claimed to need more than
+format 2 — so that build re-aims the assertion (and adds a new synthetic pair one step
+past the NEW ladder); it never re-cuts the fixture. If a truncation or hostile-name
+fixture ever changes classification, that is a regression, never a ladder consequence.
 
 ## Provenance
 
 `v1_shipping.rsbank` was produced by running this repo's own export verb (`exportBank`)
 at version **1.4.0** over a one-sample bank, and copying the emitted file here verbatim.
-Every other fixture is derived from those bytes: the truncations are prefixes of them,
-and the synthetic packages reuse their manifest region under different version integers
-or a hand-written hostile manifest (the encoder refuses to write one, which is why those
-could not come from the verb).
+Every other fixture is derived from those bytes: eight of the nine truncations are
+prefixes of `v1_shipping.rsbank` (the ninth, `trunc_additive_forward.rsbank`, is a prefix
+of `additive_forward.rsbank` itself — a prefix of a prefix, still frozen bytes, never
+regenerated), and the synthetic packages reuse their manifest region under different
+version integers or a hand-written hostile manifest (the encoder refuses to write one,
+which is why those could not come from the verb).
 
 Payloads are one 300-byte 16-bit mono WAV. The properties under test are structural —
 version integers, framing arithmetic, name validation — so a larger payload proves
@@ -46,9 +55,10 @@ build, exactly as this one was, and recording the build's version here.
 
 ### Truncation — one file per distinct decode failure site
 
-Each is a prefix of `v1_shipping.rsbank` (907-byte prefix + 300-byte payload = 1207
-bytes). All classify `Malformed`; none may classify `TooNew`, since "install a newer
-build" does not fix a partial download.
+The first eight are prefixes of `v1_shipping.rsbank` (907-byte prefix + 300-byte payload
+= 1207 bytes), so `formatVersion` never exceeds this build's on that path. All classify
+`Malformed`; none may classify `TooNew`, since "install a newer build" does not fix a
+partial download.
 
 | File | Bytes | Site the cut lands in |
 |---|---|---|
@@ -61,6 +71,13 @@ build" does not fix a partial download.
 | `trunc_payload_middle.rsbank` | 1057 | Inside the first payload. |
 | `trunc_one_short.rsbank` | 1206 | One byte short of the total. |
 
+`trunc_additive_forward.rsbank` is the ninth: a 983-byte prefix of `additive_forward.rsbank`
+(25-byte frozen header/manifest-length region + 958-byte manifest = 983), cut exactly at
+ITS payload boundary. `formatVersion` here is 2 — one past this build's — so this is the
+one truncation that proves the exact-size-proof failure stays `Malformed` even when
+`formatVersion > kPackageFormatVersion`, rather than relabeling to `TooNew` (the parse
+branch is the only one that relabels — see `src/core/package/CLAUDE.md`).
+
 ### Hostile names — refused at decode, before any planner
 
 The two naming fields carry different rules (`src/core/package/CLAUDE.md`), so each