apr run --json flag ignored — outputs human-readable text instead of JSON

[noahgift]

Bug Report

Source: tiny-model-ground-truth parity checker (0/59 passing)
Severity: Critical — blocks ALL programmatic parity checks (50/59)
apr version: 0.2.16 (0f0bcae)
Related: #239 (now fixed — inference works for SmolLM/Qwen2)

Description

After #239 fix, apr run now successfully runs inference on SmolLM-135M and Qwen2-0.5B APR models. However, the --json flag (and --format json) is completely ignored. Output is always human-readable with ANSI color codes, decorative headers, and markdown formatting — never JSON.

This blocks all programmatic parity checking since parity_check.py cannot parse the output.

Evidence

# Both --json and --format json produce identical human-readable output
$ apr run models/smollm-135m-int8.apr -p "The capital of France is" -n 32 --json 2>/dev/null

=== APR Run ===

Source: models/smollm-135m-int8.apr

Output:
\`\`\`

I would like to know if there is a way to get the capital of a country...

Completed in 5.87s (cached)

Expected JSON output:

{
  "text": "...",
  "tokens": [1234, 5678, ...],
  "token_count": 32,
  "model": "models/smollm-135m-int8.apr"
}

Note: --json IS listed in apr run --help and is accepted without error. Debug logs correctly go to stderr. The issue is only with stdout formatting.

What Needs to Happen

When --json or --format json is passed:

1. Suppress ANSI color codes and decorative headers on stdout
2. Output a JSON object with at minimum: text, tokens (token IDs), token_count
3. Keep debug/info logs on stderr only

Impact

• Blocks 50/59 parity checks (all SmolLM + Qwen2 canary, token-parity, quant-drift, roundtrip)
• Any tooling that pipes apr run output cannot parse results

Reproduction

cd tiny-model-ground-truth
make convert
apr run models/smollm-135m-int8.apr -p "test" -n 32 --json 2>/dev/null
# Expected: JSON on stdout
# Actual: Human-readable text with ANSI codes