fix(view): reverse path flags describe the published path; raw revTr loop exposed as revtr_loop_cleaned

[Burdantes]

What changed since the first version of this PR

The first version only moved a date gate (9999-12-31 → 2026-09-30) so the reverse struct's loop_detected / unresponsive_within_as would read step 04's reverse_loop / reverse_unresponsive_within_AS. Its checklist expected the reverse loop rate to drop to "a few percent"; it dropped from 75.1% to 18.3% — and investigation showed why that column cannot be published under that name:

• step 04 computes reverse_loop on the raw revTr (reverse_node_with_geo_info), before its cleanup drops ambiguous hop_type=4 excursions and truncates at the first AS re-entry (reverse_hops_clean);
• so a published reverse path never contains a loop (0 of 231,870 sampled on 2026-10-01), yet ~18% of reverse paths would show loop_detected = TRUE next to an as_path with no loop;
• and loop_detected already means "the published path" in server_to_client_path.

Now

• Both directions compute loop_detected and unresponsive_within_as from their own published hops, with one definition (A B A loop; A … A gap over unmapped hops). Reverse loop_detected is FALSE by construction and documented as such. Reverse unresponsive_within_as is now available for every date (it was NULL before 2026-09-30).
• New client_to_server_path.revtr_loop_cleaned: TRUE when the raw revTr had an AS loop that step 04 removed or truncated (the published path is then shorter than what was measured). From 2026-09-30; NULL before, since the raw path is not stored and the old definition was wrong.
• reverse_unresponsive_within_AS joins forward_loop / forward_unresponse_within_AS as physical columns the view supersedes from the hops.

Evidence

The reworked view SQL rendered as a read-only SELECT over prod partitions:

	2026-09-29	2026-09-30	2026-10-01
reverse loop_detected TRUE	0	0	0
reverse unresponsive_within_as	6.95%	6.97%	7.04%
revtr_loop_cleaned	NULL (all 2.47M)	18.26%	18.23%
forward loop_detected (unchanged)	0.89%	0.89%	0.88%

Original checklist: zeroed Welch groups 121 (09-29) → 0 (09-30, 10-01) ✅; image 4d73983 for the 10-01 nightly — inferred (built 09-30, both fixes visible from the 09-30 partition) ✅.

Follow-up (separate PR)

Recording in step 04 which cleanup happened — only a type-4 hop dropped (path complete) vs truncated (path shortened) — would make revtr_loop_cleaned more informative.

Deploy

View only: run create_events_enriched.sql with DS=hermes_union (additive: one new field in client_to_server_path).

🤖 Generated with Claude Code