O8F Report 03 — F2 live-text sourcing result (G2)

• macro: v0.6-o8f-implement-f2-f3-gap6-deploy-proof
• date_utc: 2026-05-21 · gate: G2 — implement F2 · result: G2 PASS — F2 CLOSED

1. Change (localised to cutter_agent/orchestrator/discover.py)

LiveReadOnlyDiscoverer.source_unit_texts(doc_prefix):
  NEW read-only method — SELECT iu.canonical_address, iu.sort_order,
  coalesce(uv.title,''), coalesce(uv.body,'') FROM public.information_unit iu
  JOIN public.unit_version uv ON uv.id = iu.version_anchor_ref
  AND uv.unit_id = iu.id WHERE canonical_address LIKE prefix
  ORDER BY sort_order. Returns an ordered List[(title, body)].
  Pure SELECT — the injected select_rows callable owns credentials.

LiveDryRunDiscoverer.__init__(unit_texts=...):
  NEW optional param; _shadow_mark_rows uses the real (title, body) per row
  index when present, else keeps the deterministic placeholder (fallback).
  The synthetic O4-DRYRUN-SHADOW canonical_address is KEPT regardless so a
  re-cut never collides with the already-cut live units.

LiveDryRunDiscoverer.from_live(source_live_text=True):
  Calls source_unit_texts() inside try/except — any failure (privilege,
  empty doc, fake select callable) falls back to the placeholder. No crash.

live_text_sourced_count + pin_live_context:
  Records how many shadow rows carry genuine live text; pins
  live_text_sourced_count / live_text_fully_sourced onto RunContext.

The dry-run simulator behaviour is preserved: InMemoryDiscoverer is untouched; with no unit_texts the placeholder path is byte-equivalent to the prior behaviour.

2. Tests — tests/test_orchestrator_o8f_f2_live_text.py (12 tests, all PASS)

source_unit_texts:   ordered (title,body) pairs · read-only SELECT only ·
                     empty document -> empty list
shadow rows:         real text propagates · placeholder fallback ·
                     partial (mixed real+placeholder) · shadow address kept
from_live:           sources real text · falls back when sourcing raises ·
                     source_live_text=False skips · pins recorded
cutplan phase:       cutplan_rows carry live-sourced body/title (no O4-SHADOW)

3. Real-DB read-only proof (VPS — PROOF-1)

units_sourced:          60   (= live ICX-CONST count)
units_with_real_body:   60   (genuine unit_version.body text)
units_with_real_title:  0    (live ICX-CONST rows have empty title — that
                              IS the real data; F2 sources what exists)
db_write:               NONE — read-only SELECT, transaction rolled back

4. Verdict

f2_status:  CLOSED — real unit_version body/title now sourced into the
            discover/cutplan path; placeholder retained only as fallback;
            real-DB read-only proof returned 60 genuine units.
g2:         PASS