Data-first recursive consumers in readBytes

Reorder recursive byte-stream consumers so the consumed input is inspected
before loop-control arguments can drive evaluation. Previously, partially
applying `readBytes` to a known count, such as `readBytes 2`, allowed the
evaluator to specialize the recursive worker using known counter values
while the byte stream was still abstract. This caused symbolic recursion
over unknown input and produced an enormous normal form.

The recursive worker now takes the byte stream first and immediately
case-analyzes it. As a result, partial application blocks at the input
boundary instead of unrolling the counter loop.

This preserves the fully-applied behavior of `readBytes`, while making partial
application such as `readBytes 2` normalize safely.

AGENTS.md

@@ -50,7 +50,7 @@ nix build .#
 Tests live in `test/Spec.hs` and use **Tasty** + **HUnit**.
 
 ```bash
-nix flake check   # or: nix build .#test
+nix flake check
 ```
 
 ### Test groups
@@ -91,6 +91,14 @@ head (map f xs)        → From lib/list.tri
 -- line comment
 ```
 
+CRITICAL:
+
+When working with recursion in `tricu` files:
+
+1. Put consumed data first in recursive workers.
+2. Let data shape drive recursion.
+3. Do not let counters unroll over abstract input.
+
 ## 5. Output Formats
 
 The `eval` command accepts `--form` (shorthand `-t`):