Section Citations Orphaned When Definitions Live in 1-research/

The symptom

Export a memo (HTML or PDF) and a large fraction of citation markers in the rendered output show as raw [^N] text instead of a clickable footnote, or — worse — get silently collapsed to the wrong source. On the 2026-06-14 Panthalassa-Deck-Series-B v0.0.3 export:

Body had 17 distinct citation IDs [^1] … [^17]
Only 8 unique citation definitions made it into the consolidated citation block
The fix_citations.py post-processor collapsed 173 footnote refs into 8 sources, including merging references from different sections that all happened to use [^N] locally

The root cause

The orchestrator’s writer agent emits two artifacts per section:

2-sections/NN-foo.md — the prose draft. Contains [^N] inline references.
1-research/NN-foo-research.md — the per-section research file. Contains the [^N]: … definitions and the foundational source corpus the writer was citing.

This split is documented in the manage-memo-citations skill (“foundational citations live in 1-research/, not in section files”). The pipeline relies on it.

The downstream cli/utils/consolidate_citations.py only reads ### Citations blocks inside section files. If a section file has refs in the body but no ### Citations block (because the defs live in 1-research/), the consolidator never sees the definitions. Two compounding failure modes follow:

Orphan refs. Refs in the body have no matching def in the final draft. Pandoc renders them as raw text or drops them.
Positional misalignment. consolidate_citations.py assumes content_sections[i] references citation_blocks[i] — a positional assumption that only holds if every section has a ### Citations block. When a section has refs but no block, the consolidator renumbers its refs against the next section’s def map, corrupting the bibliography.

This bug affects every memo that follows the documented research-and-section split. Until now it has been masked by per-section ### Citations blocks that some writer paths happen to emit. The Panthalassa Series B v0.0.3 export made it acute because most sections had no per-section block.

The fix

A new pre-assembly step: cli/utils/hydrate_section_citations.py.

For each 2-sections/NN-foo.md:

Collect every [^id] reference in the body.
Collect every [^id]: definition already in the section file’s ### Citations block (if any).
For each reference not yet defined locally, look up the same id in 1-research/NN-foo-research.md and pull the definition into the section file’s ### Citations block (creating the block if absent).
Report any refs that have no matching def in either place — these are real orphans (writer-invented IDs, research-file gaps) and the script prints them so the analyst can see them.

The script is idempotent. Section files that already have complete ### Citations blocks are not modified. Calling it twice in a row is a no-op the second time.

cli/recompile_memo.py was updated to invoke hydrate_section_citations.py before assemble_sections() runs. That means every existing pipeline that already calls recompile_memo.py (and every future one) gets the fix automatically. No agent prompt changes, no outline YAML changes.

Exit code 2 from the hydrator means “orphans reported” — recompile_memo.py treats that as informational, not failure. Exit code 0 means no orphans. Any other exit code is a real bug.

Effect on the Panthalassa-Deck-Series-B v0.0.3 export

Before:

Found 4 '### Citations' blocks
Consolidated to 1 citation block with 9 unique citations
…
Consolidated 173 duplicate footnotes
8 unique sources remain

After (with the new hydrator running first):

✓ 01-risks: hydrated from 1-research/
✓ 02-diligence: hydrated from 1-research/
✓ 03-capital-syndicate: hydrated from 1-research/
✓ 04-category-leadership: hydrated from 1-research/
✓ 07-colossal-market-size: hydrated from 1-research/
✓ 08-counter-cyclicality: hydrated from 1-research/
✓ 09-cash-on-cash-return-probability: hydrated from 1-research/

7 section(s) hydrated

Orphan references (referenced but never defined — will not resolve in export):
  02-diligence: [^4], [^7], [^8], [^9], [^10], [^11], [^12], [^13], [^17]
  07-colossal-market-size: [^5], [^6], [^7]

Found 10 '### Citations' blocks
Consolidated to 1 citation block with 58 unique citations
…
Consolidated 157 duplicate footnotes
47 unique sources remain

Bibliography went from 8 sources to 47. The 12 remaining orphans (in sections 02 and 07) are real — they’re refs the writer agent emitted with no matching def in either the section or the research file. They surface in the hydrator’s report so the analyst can fix them deliberately.

What the hydrator does NOT do

It does not invent definitions. If [^9] is in the section body but appears in neither the section’s existing block nor 1-research/NN-foo-research.md, it stays orphan and is reported.
It does not deduplicate across sections — that’s cli/utils/fix_citations.py’s job, run later in the export.
It does not enforce the schema of the citation definitions. Whatever the writer agent emitted is what the hydrator pulls.
It does not modify 1-research/*-research.md files. They remain the source of truth.

Anti-patterns to avoid

Don’t move section definitions out of 1-research/ to “fix” the symptom. That breaks the manage-memo-citations skill’s foundational invariant. The hydrator is the right fix; it preserves the canonical split.
Don’t run consolidate_citations.py without the hydrator step first. You’ll get the original bug. recompile_memo.py now runs both in order — use it.
Don’t hand-edit ### Citations blocks the hydrator added. They get refreshed on every assembly run; your edits will survive but the source of truth is 1-research/.