fix(repetition): skip detection for table-heavy output to prevent false positive loops

[gorevyoneticisi]

Summary

The sliding n-gram repetition detector triggers false positives on markdown table output. Structured tables naturally have high n-gram overlap because of pipe separators (|) and uniform row structure. This causes the agent to enter an infinite retry loop when generating tables with 5+ rows and 4+ columns (issue #1441).

Root Cause

detectConsecutiveRepeat() scans for tokens that repeat at a fixed period. In a markdown table, each row starts with | followed by similar cell patterns, creating exactly the kind of periodic signal the detector looks for. The detector cannot distinguish "format-driven repetition" (table rows) from "content-driven repetition" (actual loops).

Fix

Added isTableHeavy() that detects when >40% of non-empty lines are table rows (lines starting with | or separator lines). When detected, skip consecutive repeat detection entirely. This is a clean, targeted fix:

• No-op for non-table output (real loops still detected)
• No threshold tuning or complex heuristics
• Table detection is simple and fast (single pass over lines)

Changes

• packages/opencode/src/session/prompt/text-ngram-detection.ts: Added isTableHeavy() helper and gate in TextNgramMonitor.append()
• packages/opencode/test/session/text-ngram-detection.test.ts: Added 2 new tests covering streaming large tables and table-then-new-content scenarios

Test Results

All 25 n-gram detection tests pass (23 existing + 2 new). No regressions.

Closes #1441