tae-conversion: add host-side toolchain for Fenix ui/efficiency conve…

[jjSDET]

tae-conversion: host-side toolchain for Fenix ui/efficiency conversions

Branch: tae-conversion-toolchain → main
Repo: mozilla-mobile/testops-tools

What

Adds tae-conversion/ — the eff* toolchain that closes the AI-assisted conversion
loop for Fenix UI tests moving onto the ui/efficiency framework:

pick a candidate → scaffold from the legacy test → static pre-flight → build/run on a
device → read a structured verdict → file, commit, submit.

tae-conversion/
├── tools/              11 tools + .eff.env.example + the done-ledger
├── docs/               CONVERSION-LOOP, CONVERSION-LESSONS, HARNESS-GOTCHAS
└── conversion-runs/    queue protocol + prioritized candidate pool

Why here

These tools run on the engineer's machine and talk to a device, Bugzilla, and
Phabricator. The agent's sandbox can reach none of that, so effwatch.sh bridges the
two: it polls a queue directory and runs only a fixed, whitelisted command — never
arbitrary text from a request file.

That makes this infra tooling, not something to ship inside an agent plugin. The
companion PR to mozilla/firefox-aidev-plugins adds the three tae skills that drive
these tools by name.

Deliberately excluded

• effpretty and effview ship in-tree under ui/efficiency/devtools/. They're
  device-side dump tools, so in-tree is their single source of truth — duplicating them
  here would create a sync problem.
• The end-of-day project-tracker runbook. It's specific to one team's Google Sheet,
  Slack channel, and Jira tree rather than to this toolchain, and it carried a live
  Sheet ID. This repo is public. The one genuinely reusable part — refreshing the
  done-ledger with reconcile_conversion.py — is documented in the README instead.

Scrubbed for publication

Was	Now
effbug.py hardcoded a tracking epic in every filed bug's footer	Reads EFF_EPIC; omits the note when unset
effbug.py example used a real assignee address	Genericized
effsubmit.py hardcoded default reviewers	Defaults to EFF_REVIEWERS, warns when unset
effloop.sh / effwatch.sh wrote to a personal scratch directory	Default to tae-conversion/conversion-runs
Nothing ignored the Bugzilla API key file	.gitignore covers tools/.eff.env plus disposable run artifacts

Confirmed no API keys, tokens, absolute home paths, Sheet IDs, or Slack channel IDs in
the committed tree.

Freshness

Sources were taken from the live working copies, not an earlier staging snapshot that
had drifted (effloop.sh by 129 lines, effbug.py by 20, effsubmit.py by 7;
HARNESS-GOTCHAS.md by 126 and CONVERSION-LESSONS.md by 99).

The done-ledger was regenerated against current mozilla-central: 72 of 174 smoke tests
converted (67 annotated with @Converted, 5 in flight). The previously staged ledger
was four rows stale.

docs/HARNESS-GOTCHAS.md carries A1–A15 and B1–B11 and notes explicitly that the
in-tree docs/gotchas.md currently stops at A8/B8 — this copy runs ahead while entries
are still being confirmed.

Verification

• All 9 Python tools byte-compile; both shell scripts pass bash -n.
• effnext.py, effcheck.py, and reconcile_conversion.py were run from this layout
  against a real Fenix checkout. effcheck returns [PASS] statically clean on a landed
  converted test; effnext correctly reports 166 pooled / 89 done / 95 pending.
• Not exercised: effloop/effwatch/effbug/effgit/effsubmit, which need a
  connected device, a Bugzilla key, or moz-phab.