[ajhcs]

Repeated model records can exhaust an agent’s source allowance, while sampled-check input shapes and qualitative goal/evidence meanings need clearer guidance. This update adds optional lossless compact assembled context with an exact UTF-8 byte cap and explicit refusal, plus a strict scope-pinned sampled-trace adapter and discoverable schema help. The default full view and existing language checker remain unchanged.

The shared workflow now carries qualitative goal direction, distinct evidence stages, and reviews actual output omissions. It also integrates the reviewed font/SVG/reference-selection guidance, expands the adversarial corpus to 40 cases, and replaces the README with a thorough source-linked guide to all eleven skills, setup, checks, components, examples, and result limits. The technical guide’s current skill inventory is aligned.

Validation:

• Full local regression: 73 suites, 1,183 tests, zero skips, using the pinned parser and symbolic dependencies. Independent implementation reviews are closed.
• Eight focused README tests pass. Actual desktop/mobile previews load all eleven images without page overflow; source links, commands, anchors, and capability claims received independent review.
• Fresh harder comparison: 24 runs, 96 output units, 408 essential contributions. Adjudicated M=0/U=0/O=7 units, eight omitted contributions; source-complete runs A=1/6, B=6/6, C=6/6, D=4/6. All 708 fixed behavior cases pass; native monitor aggregates remain unknown.
• Two blinded initial reviews, a separate cold assessment/adjudication, process and metrics audits, and an independent report review are retained. Historical results, failed attempts, earlier grammar errors, and the separate preflight-retention/replay evidence are preserved.

The comparison report separates source completeness, finite execution, process compliance and engineering acceptance, and records the small sample, voluntary checking, common infrastructure changes, source-budget limits and remaining omissions. The dirty canonical checkout and previous results were preserved.