feat(adw): the profile is the phase list, not a label on it

[ruslan-korneev]

Stacked on #2 (which is stacked on #1). Closes RUS-190 and RUS-191 — they were filed separately, but
the phase matrix and the new profile are one edit; splitting them would have meant writing the same
table twice.

The problem

workflow-triage picked a profile and the scripts honoured it — review.sh:37 scaled the fan-out,
gate.sh:167 scaled the required lenses — but adw-run did not. Its phases were written
unconditionally, so a light run still paid for the scout agent, the full artifact set, the groom agent
passes and a stop at G2.

The two skills also disagreed in writing: triage said light → G7 only; adw-run phase 4 said
"stop for the operator" with no mention of the profile. The orchestrator wins in practice, so light
bought a cheaper groom, a cheaper review, and nothing else.

The change

adw-run carries a phase × profile table and runs one column.

Phase	superlight	light	standard / deep
scout	skip	skip	agent
plan	~12 lines + 1 check	short but complete	full
groom	none	1 lens, inline	one agent per lens
G2	skip	skip	stop
worktree + draft PR	yes	yes	yes
RED	one test, written inline	optional	agent, gate enforced
green + evidence	required	required	required
review	1 reviewer	1 reviewer	fan-out + wildcard (+ judge)
G7	stop	stop	stop

Below standard the specialist agents go away, because their value goes away: the scout protects a
context this task will not fill, and the test-author guard exists so an implementer cannot bend a test —
when you are about to write both, the honest substitute is writing the test first and not touching it
after. G2 goes away because a twelve-line plan is one the operator approves without reading, and a gate
that gets rubber-stamped devalues the one that must not.

superlight is new: no groom pass at all, one check, one reviewer. It needed no change to
gate.sh's requirements — gate.sh plan wants four headers and one numbered check, gate.sh evidence
wants one evidence file, and on a task this size three of those headers are one word long.

What never scales down: the worktree, the draft PR, gate.sh green, gate.sh evidence, one
reviewer. The workspace takes seconds and is what makes a run revertible and visible; one reviewer is
the cheapest independent look at a diff there is.

The floor, which nothing documented before: a typo, a version bump, a comment fix should not enter
the loop at all. Triage reasons about blast radius, so it honestly answers superlight for a one-liner
— that is not a licence to spend a worktree and a PR on it.

Operator override

arguments: [profile, ref]
argument-hint: "[superlight|light|standard|deep] [ticket-ref]"

/adw-run light SM-12. Claude Code's : is a plugin namespace (/my-plugin:review), not an option
separator — /adw-run:light would parse as a skill light in a plugin adw-run. Arguments are the
mechanism, and an unpassed named argument expands to an empty string, so "nobody chose" stays a normal
branch.

Script changes

• gate.sh: superlight added to the profile alternation. Without it, **Profile:** superlight matches
  the light inside the word and the run is groomed as light. superlight also skips the GROOM_LOG.md
  requirement — no pass was meant to happen — while the blocker check still runs, because a blocker stops
  the run at every profile.
• review.sh: superlight resolves to the light fan-out instead of dying as an unknown profile.

How to verify

bash tests/run.sh                            # 116 passed, 0 failed, 1 known gap
ADW_TEST_SHELL=/bin/bash bash tests/run.sh

Non-vacuous by mutation: dropping superlight from gate.sh's alternation fails 2 assertions, dropping
it from review.sh fails 1.

One thing worth knowing, because a comment in the first draft of this PR got it wrong: order in the
alternation is irrelevant. ERE is leftmost-longest, so light|superlight and superlight|light both
return superlight — verified by swapping them and watching the suite stay green. What matters is that
the longer word is in the list at all.

Not verified

• No real task has been run at superlight end to end. The profile is asserted at the two places a
  script can see it (gate.sh groom, review.sh); the rest is instructions to the orchestrator, and
  instructions are advisory by construction.
• The floor is prose. Nothing mechanically stops a one-line change from entering the loop, and nothing
  should — the judgement is the operator's.
• arguments:/argument-hint: frontmatter is documented Claude Code behaviour but has not been
  exercised here; /adw-run light SM-12 has not been typed against the installed skill.