feat(autofix): audit the approach instead of stopping on growth-budget breach

.github/workflows/qwen-autofix.md

@@ -93,8 +93,6 @@ YAML, and never delete a section without deleting its pointer.
 - [43. review-address · Prepare branch and feedback — Growth brake: measure the PR's net size (insertions minus deletions vs the merge base),…](#af-043)
 - [44. review-address · Prepare branch and feedback — An orphan-history branch (fork takeover / adoption admits one — nothing on this job's…](#af-044)
 - [45. review-address · Prepare branch and feedback — The marker's window field is spelled `key=`, NOT `win=`: this marker can legitimately…](#af-045)
-- [46. review-address · Prepare branch and feedback — Divergence: Critical-only only trims non-Criticals, so when the GROWTH that trips the…](#af-046)
-- [47. review-address · Prepare branch and feedback — Count runs whenever the net is measured (not only over budget), so the trajectory clause…](#af-047)
 - [48. review-address · Prepare branch and feedback — Which trusted humans have exhausted their per-window regular feedback budget (see…](#af-048)
 - [49. review-address · Prepare branch and feedback — Time-budget exhaustions SINCE THE LAST SUCCESSFUL ROUND mean the standard…](#af-049)
 - [50. review-address · Triage and address — Bound the agent below the job timeout so a runaway agent fails THIS step (not the whole…](#af-050)
@@ -1397,87 +1395,6 @@ post-update size. (A conflict round's own merge of main is the
 narrower residual; its delta is bounded by the overlap.)
 ```
 
-<a id="af-046"></a>
-
-### 46. review-address · Prepare branch and feedback — Divergence: Critical-only only trims non-Criticals, so when the GROWTH that trips the…
-
-In `review-address` · `Prepare branch and feedback`.
-
-```text
-Divergence: Critical-only only trims non-Criticals, so when the
-GROWTH that trips the brake is Critical-driven the diff keeps
-climbing anyway. Read this window's prior per-round growth markers
-(written by the report step): count the rounds that were over
-budget, and take the MOST RECENT prior over-budget run's growth
-SUM (latest measured= — see below; NOT the window-wide max, which a
-one-off spike would raise forever). The round is DIVERGING when it
-is over budget now, the brake has already fired for
->= GROWTH_DIVERGENCE_ROUNDS prior rounds, and the diff has NOT
-shrunk from that most-recent sum — the fixes are not converging, so
-the round must escalate to a human decision instead of patching
-again. A diff that is over budget but SHRINKING (agent removing
-code) or a one-off overshoot stays in ordinary Critical-only.
-```
-
-<a id="af-047"></a>
-
-### 47. review-address · Prepare branch and feedback — Count runs whenever the net is measured (not only over budget), so the trajectory clause…
-
-In `review-address` · `Prepare branch and feedback`.
-
-```text
-Count runs whenever the net is measured (not only over budget), so
-the trajectory clause below is accurate even on a round that pulled
-back under budget. markers:
-<!-- autofix-growth-now src=N test=N over=BOOL round=N run=ID measured=TS key=W -->
-Deduped by run=GITHUB_RUN_ID (the per-workflow-run id) and ORDERED
-by measured=: the report post's bounded retry re-posts one run's
-marker, and a failed job's re-run keeps the same run_id, so a run
-collapses to its LATEST measurement — and that collapse happens
-BEFORE the over/window/cutoff filters, or a re-run that came back
-under budget would still be represented by its stale over=true
-attempt. Within the collapse an explicit measured= beats the
-created_at fallback: a re-run attempt that crashed BEFORE prepare
-— or whose measurement failed — posts an inert over=false marker
-with no measured=, whose fallback (post-run) timestamp would
-otherwise outdate and erase the same run's real prepare-time
-measurement. Every distinct address run has a fresh run_id.
-KNOWN RESIDUAL (#9114): during the one-time deploy transition a
-run whose FIRST attempt posted a legacy (no measured=) over=true
-marker and whose re-run crashes before prepare still collapses
-fallback-vs-fallback on created_at — the later inert marker wins
-and erases the count. Self-limiting: once deployed, every real
-measurement carries measured= and beats any inert marker.
-round=/eval-watermark are NOT a safe identity — a state-triggered
-lane (a persistent merge conflict selects the PR every scan with no
-new evaluable feedback) freezes both NEWEST and ROUND, so distinct
-over-budget runs would share them and collapse, stalling the count.
-Filtered on measured= (the prepare-time measurement instant, NOT
-the comment's post-agent created_at) after GROWTH_NOW_CUTOFF, so a
-prior sum measured against a pre-base-update tree is dropped rather
-than compared to this round's. KNOWN RESIDUAL (#9114): the tree is
-fixed at the branch fetch/checkout while the cutoff comes from
-ic.json fetched afterwards, so a base update landing between the
-fetch and the measured_at stamp admits a pre-update marker;
-self-heals at the next re-arm/base update. measured= is OPTIONAL in
-the scan:
-markers posted before it existed fall back to their comment's
-created_at, so deploying this does not blank the census of a window
-that is already in flight. KNOWN RESIDUAL (#9114): during that
-transition the sort mixes two clocks — a legacy marker's fallback
-is its POST-RUN created_at while a new marker stamps prepare time —
-so PREV_SUM can briefly come from an older measurement; the count
-is unaffected and it self-heals at the next re-arm/base update.
-The "not shrinking" test compares against the MOST RECENT prior
-over-budget run's sum (latest measured=), not the window-wide max: a
-single transient spike would otherwise raise the bar forever and a
-genuine plateau-over-budget runaway (the exact case to escalate)
-would never clear it. The CURRENT run's own markers are excluded
-(run != GITHUB_RUN_ID): a re-run of a failed job keeps the same run
-id and its failed attempt already posted a marker, so counting it
-would over-report the round's own attempt as a PRIOR one.
-```
-
 <a id="af-048"></a>
 
 ### 48. review-address · Prepare branch and feedback — Which trusted humans have exhausted their per-window regular feedback budget (see…