Fix advice budget and retain native release test diagnostics - #8

Problem and change

The 0.14.8 release runtime gate found that a 120-character advice request returned a 229-character block. Advice now counts the JSON-quoted query, headings, and truncation marker in the overall budget; budgets that cannot retain a complete row marker return no block. Renderer-owned row boundaries keep feedback IDs limited to retained rows, even when memory content quotes another row marker. Recall captures stay together across suspension.

The native-test gate also retains output and captures bounded process stacks after prolonged silence, so a stalled run provides diagnostic evidence. Its Swift command, exclusions, success requirement, and 30-minute runtime budget remain unchanged.

Validation

• Existing end-to-end advice contract is unchanged; focused Swift boundary and feedback-ID regressions added.
• Swift syntax parsing and diff checks passed; native execution remains required in the release gate.
• Twelve focused Python supervisor tests passed, covering timeout/cancellation, bounded diagnostics, output draining, process cleanup, and exit propagation. Three release workflow fixture tests passed, including exact Swift argument verification.
• Failed release attempt 35441295700 retained: fresh Core2.0.5 build passed,507 test functions passed, one advice contract failed, four unfinished when cancelled. No release was published.