Bundles the extraction perf work + the linking-audit fixes into one PR.

Volume / cost

• prospect.question signal-gated , capture only when it reveals a need, carries an objection/hesitation, probes the buying process, or triggers a demo. It was 25% of all rows and the driver of the group-demo output-cap blowout.
• Slim vs full envelope , count-and-classify fields (prospect.question, next_step.set, stakeholder.engaged, tactic.*, prospect.call_ran_over) emit only content+payload+persona+ids (null summary/responseSpan/strength/completeness). Measured: prospect.question −43%/moment. Enforced defensively in assemble_sonnet.py.
• Per-plane defs , gen-field-defs emits <stem>.instance.ts (instance + call-level) and <stem>.judgment.ts (versioned); each agent reads only its plane. Judgment 115KB→27KB (−77%), instance →90KB (−22%). Net def-read load per 3-meeting deal ~196k→~72k tokens.
• Classify merged into instance , the call is classified (call_type/demo_story_type) first in the same transcript read; the separate classify fan-out is gone (one fewer agent + one fewer transcript read per meeting).
• Output-cap retry , a failed instance agent retries once with a stricter signal budget (Sonnet's 64k output is a hard model ceiling, not a raisable env), so a dense call isn't silently lost.

CRM-leak fix (was missing on this branch , overlaps open #2506)

gen-field-defs now filters background:true first, so an org's live CRM list/text fields (50 for Pirros) no longer leak into the extraction planes , Pirros went from a wrong 29 instance + 16 versioned back to the canonical 18 + 8. Without this we'd extract transcript values into live CRM fields. Vocab guardrail + multi-select sidecar scoped to the taxonomy plane too. Note: this duplicates #2506's fix , recommend closing #2506 or merging this and rebasing.

Linking correctness (from the linking audit)

• SEV-1a , pointer refs are the target atom's local_id / prior base#local id, not a "verbatim span" (the resolver drops span refs as danglers). Fixed in the extract prompt and the 5 shipped field defs (responds_to, callback_to, tie_back×2, pains_covered).
• SEV-1b , links are driven by the relationship, with the companion label (flow=callback / narrative_type⊇pain_bridge / mode=recap) set to agree , no longer "set ONLY when <label>", which suppressed recall (0 of 52 links in the dry run). Added a dedicated linking pass.
• SEV-2a , write-agentic builds cross-call priorCandidates from the DB (the deal's prior meetings' pain/use_case/rep.question atoms) and event-scopes the delete, so cross-call refs resolve on partial/incremental runs instead of dangling. Full-run behavior is byte-identical.

Safety / tests

• Field-def changes are description-only (field set/order unchanged). Equivalence content hash re-pinned; defaults-equivalence + link-resolver tests green (12/12). eval-new-fields.ts regenerated.
• Typecheck clean on all touched files (pre-existing errors elsewhere untouched).
• All DB writes remain behind write-agentic's DRY_RUN gate , recommend a --dry-run pass before the next real write to confirm cross-call priorCandidates resolution.

🤖 Generated with Claude Code

https://claude.ai/code/session_018Tggg2sBmoV3cCZGMvpsdX

Greptile Summary

The PR reduces taxonomy-evaluation token volume, splits generated definitions by extraction plane, merges call classification into instance extraction, and improves cross-call pointer resolution. It also changes incremental write replacement and retry behavior, where two correctness issues remain.

• Signal-gates and slims high-volume instance fields.
• Generates dedicated instance and judgment definition modules.
• Retries failed instance extraction once under a stricter output budget.
• Filt