pipeline: reject resref collisions when chunking; document erf post-write hashing (#8)

## What

Two correctness fixes surfaced in review of the direct-depot HAK artifact work.

### Reject resref+type collisions in `chunksFromManifest`

A manifest can name two distinct source paths that collapse to the same
`resref+type` (e.g. `creature/foo.tga` and `placeable/foo.tga` — resref is the
lowercase basename minus extension). The ERF writer keys resources on
`Name:Type`, so the second silently shadowed the first: an asset would vanish
from the packed HAK with no error.

`chunksFromManifest` now runs `ensureUniqueChunkResources` per chunk and fails
the build on a duplicate. This guards **both** build paths — the legacy
`--source-manifest` flow and the direct content-addressed flow
(`chunksFromSourceManifest` → `chunksFromManifest`).

### Document erf post-write hash coupling

`writeResourceData` streams the source into the output while hashing, so
size/SHA mismatches are only detected *after* the bytes are written. A non-nil
return therefore means the writer holds partial, unverified output and the
caller must discard it. Added a comment making that contract explicit;
`writeHAKArchive` already honours it (writes to a temp file, removes on any
Write error, never renames a bad archive into place).

## Tests

- `TestChunksFromManifestRejectsResrefCollision`: collision → error, distinct
  resrefs → clean. Asserts only error presence/absence — silent asset loss is
  the contract, not any specific wording.
- `go vet ./internal/erf/ ./internal/pipeline/` clean.
- `go test ./internal/erf/ ./internal/pipeline/` green.

## Follow-up

A new crucible release must be cut after this merges so the guard ships in the
binary `sow-assets-manifest` pins.

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

Reviewed-on: #8
Co-authored-by: vickydotbat <vickydotbat@tutamail.com>
Co-committed-by: vickydotbat <vickydotbat@tutamail.com>

AGENTS.md

@@ -42,3 +42,7 @@ make build                  # cmd/* -> ./bin
 make smoke                  # assert fail-closed contract
 make image                  # crucible:<sha>
 ```
+
+## Tests
+
+Tests must survive harmless changes to constants, defaults, wording, ordering, fixture data, and internal implementation details. A test that fails merely because a basic value changed is usually a bad test. Only assert exact values when the value is part of a documented public contract, external protocol, compatibility requirement, security rule, migration, or business rule.