Messaging and calendar screen categories now emit strict, bounded JSON with category formatter modules that render dict outputs back into the familiar transcript shape. Category discovery also carries a per-category output-token cap; categories without a frontmatter override still use 4096. D3 token derivation: - Real-data basis (n=1304 content.messaging frames, real journal): total chars median 1055 / p90 2522 / p99 4389 / max 7303; messages per frame max 41; longest single message text 2859 chars; sender max 64; timestamp max 65. - Bounds sit ABOVE observed maxima on purpose: messages maxItems 60 (>41), text maxLength 4000 (>2859), sender 128 (>64), timestamp 96 (>65). schema_prep.py strips maxItems/maxLength from Anthropic requests, but models.py still validates the response against the canonical bounded schema, so a bound sized at the median is a production SchemaValidationError, not a lint failure. - Output cap: a faithful worst-case extraction re-encoded as JSON is ~7303 chars of content plus key/quote/escape overhead (~1.45x), about 10.6k chars, or about 3.0k tokens at 3.5 chars/token. The previous global cap of 4096 output tokens (~14.3k chars) left about 26% headroom over the observed p100, and JSON truncation is fatal (JSONDecodeError -> retry -> error) where markdown truncation degraded gracefully. Messaging and calendar therefore get max_output_tokens: 8192 via new category frontmatter while the default stays 4096: about 2.7x the observed p100 and about 6x p99. - Explicitly rejected: proving the schema's theoretical maximum serialization fits 4096 tokens. For an array-of-free-text schema that ceiling is maxItems x maxLength (60 x 4000 = 240k chars) and is not a useful bound; forcing it under 14.3k chars would cap text near 200 chars and truncate real 2859-char messages. Hence the per-category frontmatter key rather than a tighter schema. - Side effect, intended: both categories now take the JSON thinking_budget of 6144 (was 4096) via the existing 6144 if is_json else 4096 path in describe.py. D4 no migration: - 1304 on-disk frames keep content.messaging as a markdown str. Both formatters return "" for non-dict input, routing legacy frames through the existing str fallback in screen.py::_format_category_content (**Messaging:** plus the raw body), byte-identical to main. This is the meeting.py contract, not a compat shim. No migration script, no data rewrite. Coverage honesty: - calendar has ZERO real frames in the journal; calendar.md was added in 97a97a7b and has never produced an extraction. Its coverage is synthetic fixtures only. - make eval-schemas was NOT run: no local provider is configured, so tests/eval_schemas.py raises LOCAL_NOT_READY. No qwen results are claimed. scripts/check_schema_bounds.py::ALLOWLIST is unchanged. The four talent schemas labelled messaging/calendar follow-on lode (chat, participation, participation_entry, schedule) are talent schemas, out of scope for this lode, and remain allowlisted. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>