fix(run): raise review findings JSON cap 5000→20000 chars

[johnkattenhorn]

Problem

In run_review_loop (ralph/ralph_loop.sh:1957, 1963) the extracted JSON between the ---REVIEW_FINDINGS--- and ---END_REVIEW_FINDINGS--- markers is truncated to 5000 chars before being passed to jq for validation.

findings_json=$(sed -n '/---REVIEW_FINDINGS---/,/---END_REVIEW_FINDINGS---/{//!p;}' "$review_output_file" 2>/dev/null | tr -d '\n' | head -c 5000)

Any review that produces more than 5000 chars of JSON — which is trivial to hit with a project-specific REVIEW_PROMPT.md that asks for file, line, category, issue, and a substantive suggestion per detail entry — gets cut mid-object. jq then rejects it as malformed syntax and the whole batch of findings is silently dropped:

[2026-04-17 12:18:21] [WARN] Review findings JSON is malformed — skipping

Ralph never sees the findings. The review loop burns full tokens but provides no signal to the next implementation loop.

Repro

Observed in a real project running REVIEW_MODE=ultimate with a customised REVIEW_PROMPT.md. Six reviews ran:

Review	JSON size	Outcome
Loop 2	2991 chars	✅ parsed
Loop 4	5500 chars	❌ malformed (truncated at 5000)
Loop 5	5931 chars	❌ malformed
Loop 6	5546 chars	❌ malformed
Loop 7	5073 chars	❌ malformed (just over)
Loop 8	4162 chars	✅ parsed

Every review over 5000 chars failed. The lost findings included real HIGH-severity items the reviewer caught (missing cross-user isolation tests, pattern drift, architectural deviations) that Ralph would otherwise have addressed in the following loop.

Fix

Raise both head -c 5000 occurrences to head -c 20000. That's comfortable headroom for ~8–12 detailed findings per review while preserving the cap as a safety rail against a runaway model response.

No behaviour change for smaller reviews.

Test plan

• Manual: re-ran a previously-failing review_loop output file through the parser with the new limit — JSON now parses cleanly, jq .issues_found returns the correct count.
• No existing tests hit the truncation path (all test fixtures are small), so no test updates needed.
• Would welcome a suggestion on where a bats test guarding this would belong — tests/bash/ structure is mature but I didn't want to speculate.

Related

Discovered alongside #165 (prepare-script fix) while running a real project against a custom REVIEW_PROMPT.md. Both are independent.