DEVEX-1653: flip Slack classifier to bright-line on_rt rule (Phase 3)

[pdodgen-revparts]

Context

The DEVEX-1653 sweep completed 2026-07-09:

• DEVEX-1653: declare on_rt input on slack-deploy-notification (unblock sweep) #148 — added on_rt input to slack-deploy-notification.yaml (accepted, not yet consumed). Merged as ab8282c8.
• 21 sweep PRs — threaded on_rt through every RT participant's deploy chain. All merged (webstore #4670, catalog_api #650, ... — see DEVEX-1653: declare on_rt input on slack-deploy-notification (unblock sweep) #148 body for the list).

This PR is Phase 3: flip the classifier from tag-shape inference to bright-line rule based on the declared on_rt intent.

Change

- if [[ "$IMAGE_TAG" =~ -rc\.[0-9]{4}-[0-9]{2}-[0-9]{2}\.[0-9]+$ ]]; then
+ if [[ "$ON_RT" == "true" ]]; then
    BADGE_MRKDWN="🚂 *RT*"
    ...
  else
    BADGE_MRKDWN="⚡ *Hotfix*"
    ...
  fi

Plus updated env var (ON_RT: ${{ inputs.on_rt }} instead of IMAGE_TAG) and the block-comment rationale reflecting the new logic.

Why the flip

Tag inference could match v0.440.10-rc.2026-07-09.8 as 🚂 RT even when it was actually a hotfix backport manually promoted after Thursday's train ship (Sam's case). The dispatcher's intent is the source of truth, not the tag pattern.

Companion PR

• encodium/actions#TBD — rt-promote.sh update to dispatch with -f on_rt=true. Also opened as a draft, in this same window.

Merge order safety

Either order works — the on_rt input is already accepted by the shared workflow. Between merges:

Merge state	Classifier reads	Legit train ship rendered as
Neither	IMAGE_TAG (tag inference)	🚂 RT (correct — rc.1 matches pattern)
Only rt-promote.sh update	IMAGE_TAG (tag inference)	🚂 RT (still correct)
Only this PR	ON_RT (bright-line)	⚡ Hotfix (regression — under-badged)
Both	ON_RT (bright-line)	🚂 RT (correct — rt-promote.sh sets true)

Merge the rt-promote.sh PR first if there's a next promote in flight; otherwise either order is fine.

[cursor]

PR Summary

Low Risk
Observability-only change to Slack badge text; brief mis-labeling is possible if this lands before rt-promote.sh passes on_rt=true for train promotes.

Overview
Slack prod deploy notifications now label 🚂 RT vs ⚡ Hotfix from the workflow input on_rt instead of inferring from the image_tag regex (-rc.YYYY-MM-DD.N).

Only dispatches that set on_rt=true (intended for rt-promote.sh) get the RT badge; everything else defaults to Hotfix. Comments on the input and classify step were updated to document the bright-line rule and why tag shape was wrong for manual rc.N>1 promotes.

The classify step env switches from IMAGE_TAG to ON_RT, and the log line reports on_rt=… → classified as ….

^{Reviewed by Cursor Bugbot for commit f33d6cf. Bugbot is set up for automated code reviews on this repo. Configure here.}