docs: log fuzzy matching and entity count issues for Phase 8

Author: Siddharth-Khattar

.planning/MINOR_ISSUES.md

@@ -20,3 +20,26 @@ Deferred issues that are non-blocking but should be addressed in future work.
 - **Fix**: Either annotate concrete functions to return `tuple[DomainAgentOutput | None, UUID | None]`, or use a generic `TypeVar` on the return type.
 - **Priority**: Low (no runtime impact, passes pyright default mode)
 - **Added**: 2025-02-07, QC review of Phase 7 execution ID fix
+
+## MI-003: Fuzzy entity matching produces false positives on numeric/temporal values
+
+- **File**: `backend/app/services/kg_builder.py`, `deduplicate_entities()` (~line 270-288)
+- **Detail**: Fuzzy string matching (rapidfuzz ratio >=85%) flags semantically distinct values as potential duplicates when they are string-similar but numerically different. Observed in live testing:
+  - `'2016-05-02 20:00'` vs `'2016-05-02 22:00'` (92%) — different timestamps (8 PM vs 10 PM)
+  - `'50,000 dollars'` vs `'2,000 dollars'` (88%) — 25x difference in amount
+  - `'$5,000'` vs `'$50,000'` (88%) — 10x difference in amount
+- **Impact**: No data corruption (flags only, not auto-merged). But these false flags add noise for Phase 8 LLM resolution, wasting tokens and potentially confusing the synthesis agent.
+- **Fix**: Add type-aware matching logic before fuzzy comparison. For `entity_type` in (`timestamp`, `monetary_amount`, `date`, `other` when value looks numeric): parse the actual value and compare semantically instead of string-matching. E.g., for monetary amounts, extract the number and compare magnitude; for timestamps, parse and compare actual time difference.
+- **Priority**: Medium (should be fixed before or during Phase 8 to avoid noisy LLM resolution input)
+- **Phase**: Fix during Phase 8 (Synthesis) when fuzzy flags are consumed
+- **Added**: 2026-02-07, live pipeline testing
+
+## MI-004: Pipeline summary log mixes two different entity count semantics
+
+- **File**: `backend/app/services/pipeline.py` line ~1089
+- **Log**: `entities=40 kg_entities=61` — confusing because both sound like entity counts but measure different things.
+- **Detail**: `entities` = triage file-level entities (line 1048: `sum(len(fr.entities) for fr in triage_output.file_results)`) + domain agent top-level `output.entities` (line 1043). `kg_entities` = KG Builder total (includes both top-level AND per-finding entities). The KG count is always >= domain entity count because KG Builder also extracts from `finding.entities` lists. The naming makes them appear comparable when they are not.
+- **Fix**: Rename `entities` to `triage_entities` and `total_domain_entities` to `domain_output_entities` in the log, or remove the triage entity count from this summary (it's already logged in the triage stage). The `processing-complete` SSE event (line 1065) also sums these two, which is sent to the frontend — verify frontend doesn't display this misleadingly.
+- **Priority**: Low (cosmetic log clarity, no data impact)
+- **Phase**: Fix opportunistically during Phase 8 pipeline extension
+- **Added**: 2026-02-07, live pipeline testing

.planning/ROADMAP.md

@@ -830,6 +830,10 @@ Plans:
   - Task deduplication via existing task list injection into synthesis prompt
   - Stored in investigation_tasks table
 
+**Known Issues to Resolve (from `.planning/MINOR_ISSUES.md`):**
+- **MI-003**: Fuzzy entity deduplication produces false positives on numeric/temporal values (e.g., `$5,000` vs `$50,000` flagged at 88% similarity). Fix: add type-aware matching in `kg_builder.py:deduplicate_entities()` before fuzzy comparison — parse numeric entity types semantically instead of string-matching. Must be fixed before synthesis consumes fuzzy flags.
+- **MI-004**: Pipeline summary log mixes triage entity count + domain entity count as `entities=N`, confusingly alongside `kg_entities=M`. Fix: rename or separate the counters in `pipeline.py` for clarity.
+
 **Technical Notes:**
 - Synthesis Agent runs in fresh stage-isolated ADK session (consistent with existing pattern)
 - Input is TEXT from PostgreSQL (case_findings.finding_text), NOT multimodal file content