Final PR (3 of 3) closing #71. Depends on:

• TechDevGroup/devagentic#220 (graftedContextById query) — must merge first
• feat(mcp): grafted_context_fetch tool — extends devagentic-lane-h plugin (#71 PR2) #72 (grafted_context_fetch MCP tool) — must merge first

Summary

G1's preamble loader no longer inlines graft bodies. Each grafted-context becomes a single index line: - `graft_id` · source · `path` — abstract (first 120 chars of content). Worker fetches full bodies on demand via the grafted_context_fetch(graft_id) MCP tool added in PR2.

Loader explicitly teaches the worker the idiom in the preamble itself:

Full bodies are NOT included in this preamble. Use the grafted_context_fetch(graft_id) MCP tool to load any row's full content on demand — typical pattern is to skim the index, identify relevant rows by source/path/abstract, then fetch only what you need for the current turn.

Per-turn cost reduction

Old caps removed: _PER_GRAFT_CONTENT_CHARS=2048 and _MAX_GRAFTS_RENDERED=8 are deleted (no longer meaningful — content isn't in the preamble at all). New cap: _ABSTRACT_CHARS=120 per graft. _MAX_TOTAL_PREAMBLE_CHARS=32768 remains as a safety net.

What stays inlined

• Vertical-spec header (name, user_id, created_ts, manifest path, source count) — small, always relevant.
• Worker-guardrails — short, critical, can't reasonably be lazy-loaded since they govern every move the worker makes.

Test plan

• 11 tests in tests/plugins/devagentic-vertical-preamble/test_preamble.py cover: no-user_id gate, default-profile gate, empty-rollup gate, NEW: index-not-full-body assertion (poly-explorer-scale fixture with 7KB content body verifies full body NOT inlined, only abstract), NEW: 70-graft all-indexed assertion, NEW: abstract-truncated-to-120-chars assertion, load-once gate, fetch exception, resolve_user_id exception, client env override, client empty-user_id guard.
• All 11 pass.
• Sibling tests (test_grafted_context_fetch.py from PR2, test_client.py from G4): no regression.

Polynomial-explorer end-to-end after deploy

1. Devbox restart picks up devagentic#220 (server side query).
2. Container rebuild picks up feat(mcp): grafted_context_fetch tool — extends devagentic-lane-h plugin (#71 PR2) #72 + this PR (hermes side: MCP tool + index loader).
3. Next polynomial-explorer turn:
   • Preamble: ~11KB index + ~1KB guardrails + vertical-spec header ≈ 12KB (was 16-32KB).
   • Worker sees 70 grafts in the index. Reads abstracts. Identifies relevant rows for the current concept-permutation step.
   • Fetches selectively via grafted_context_fetch("doc-graft-...") — one or two at a time.
   • Model attention now scales with what the worker asks for, not what the vertical has.

Empty-content failure (the symptom that surfaced #71) should resolve organically — model no longer overwhelmed by tools+preamble+grafts. PR #69's synthetic recovery remains as a safety net for any residual failures.

Notes

• Index ordering is preserved from the rollup (path-sorted server-side by verticalContext).
• One-line abstract format chosen for grep-ability: - `<id>` · <source> · `<path>` — <abstract>. Workers can grep the preamble for source/path keywords to find candidate grafts.
• No "showing first N" suffix anymore — ALL grafts are in the index since each line is ~150 chars.