P2: Deepen prompt-ingest segmentation & distillation (ADR 0006 follow-ups)

[macanderson]

Follow-up to #44 / ADR 0006. The prompt-ingest provider (contextgraph-host/src/ingest.rs) ships a deliberately small, deterministic segmenter behind a single classify seam. The frame shapes are final and conformance-covered; these are additive quality improvements that never touch the wire.

Segmentation

• Stack traces as a first-class kind. Today a stack trace folds into Log detection (the STACK_MARKERS heuristic). A dedicated StackTrace kind with its own distiller — keep the exception/panic line + top N frames + … (M more frames) — would beat the generic log distiller for the common "paste a traceback" case.
• Richer tabular ingestion. looks_like_table / distill_table handle pipe and tab delimiters only. Add CSV (comma) and whitespace-aligned columns, and improve infer_column_type (currency, percent, nullable columns).
• Classification precedence edge cases. Table detection runs before log, so a pipe-delimited log (ts | LEVEL | msg) classifies as Fact rather than Episode; a scheme-less example.com/path token becomes an anchor. Both stay honest (cost/invariants hold), so they're documented limits — worth refining the precedence and looks_like_path precision.

Distillation

• Normalize common log timestamps to the F4 profile. leading_timestamp only sets valid_from/valid_to when the first token is already exactly YYYY-MM-DDTHH:MM:SSZ. Real logs (2026-07-20 18:00:01, syslog, bracketed) never match, so the temporal window rarely populates. Parse a few common shapes and normalize to F4 (guarded — never emit a non-F4 string, per Close enforcement gaps: audit max_frames, validate temporal fields as RFC 3339, probe as_of #10).
• Duplicate-line collapse in logs. Collapse runs of identical log lines to … (×K) before selecting salient lines, so a spammy retry loop doesn't dominate the compact rendering.

Scope note: distillation quality is provider policy (ADR 0006), so none of this is normative — it only changes what the prompt-ingest provider chooses to inline.

[macanderson]

Re-verified 2026-07-26 against origin/main @ 1ee9b1c.

ingest.rs has exactly one commit and is unchanged on origin/main. Every seam the body names verifies line-for-line: STACK_MARKERS folds tracebacks into Log with no StackTrace kind, table detection branches only on pipe then tab, and no duplicate-line collapse exists anywhere.

Recommended scope: keep as-is — no edits needed. Every claim in the body still holds exactly as written.