The test conftest cannot resolve the repo root without .git, and the marker turns that into a false 'not vendored' claim

[topij]

scripts/tests/conftest.py resolves the repo root with
_repo_layout.find_repo_root, which walks up for a .git marker and — finding
none — falls back to start.parent. That fallback is:

• correct when the engines sit directly under the root (scripts/), and
• one level short when they are nested (scripts/devkit/), the layout
  /adopt defaults to.

Nothing distinguishes the two, because the only signal that would (.git) is
exactly what is absent. This is #60's defect, already pinned by
test_repo_layout.py::test_the_fallback_is_wrong_in_a_nested_layout_and_that_is_known.

Why it needed handling in the test conftest specifically

PR #232 added a kit_repo_only marker that skips a test when a path it
needs is absent. That turns a wrong root from a loud failure into a statement:
not vendored in this tree, about a file that may be sitting right there.

Two attempts, and what each cost — recorded so a third does not repeat them:

1. Round 1: disable skipping entirely when no .git is found. Preserves the
   loud failure in the nested case. But the fallback root is correct for the
   flat layout, so this broke a case that worked: a tarball export of a
   genuinely sized-down tree went from an accurate skip to a failure — Kit tests hardcode parents[2] as the repo root, so they fail or error in the scripts/devkit/ layout /adopt defaults to #134's
   own harm class, for a different population. Withdrawn.

2. Round 2: search both REPO_ROOT and its parent. Withdrawn in round 3,
   for two reasons found by execution. It is still wrong at nesting depth > 1 —
   a tree with engines at tools/internal/devkit/ and init.sh at the true
   root got a confident not vendored in this tree about a file that was right
   there, which is worse than round 1's behaviour on the same tree. And in the
   flat layout the second candidate sits outside the repo entirely, so a
   same-named file in the parent directory — a tarball unpacked inside another
   checkout — suppressed a skip that should have fired.

3. Round 3 (shipped): no guess at all. One root is searched, and when no
   .git was found the skip reason says so:

   not vendored in this tree: init.sh (repo root unresolved — no .git above <dir>; see #233)
   

   This makes no false claim — it states exactly what was searched — and keeps
   the clean run a sized-down adopter is owed. The cost is that in a nested
   no-.git tree a test may skip when it could have run; the reason text is
   what makes that legible rather than misleading.

Three consecutive review rounds each found a defect in the previous round's
guess. That is the signal, not any single defect: the root is not derivable
from the information available, and every approximation has leaked in a
different tree shape. A fourth approximation should not be attempted here.

What a resolution needs

Something that identifies the repo root without .git. Options, none costed:

• Read paths.engines from config/dev-model.yaml and walk up that many
  segments from the engines directory. The config is the value being hardcoded
  today; it is also the thing the layout is defined by. Cost: the conftest
  gains a YAML read, and kitconfig is not importable at conftest time without
  the same path arithmetic being solved first.
• A repo-root sentinel the kit writes and adopters keep.
• Accept the widening permanently and document it as the answer.

Related: #60 (the resolver limit itself), #203 (three modules still carry
private copies of the walk-up), #134, #232.

[topij]

Measured: a no-.git tree still aborts collection, for a reason predating the marker.

Built the by-the-book /adopt tree at PR #232's round-3 head and removed .git
(a git archive export before git init, the case this issue is about):

!!!!!! Interrupted: 3 errors during collection !!!!!!
E   RuntimeError: no repository root above .../tree/scripts/devkit

Three modules — test_portability.py, test_mutation_gate.py and
test_pr_followup_hook.py — carry private walk-up helpers that raise
rather than falling back, which _repo_layout.py's own docstring names:

test_portability.py's private copy raises there instead; raising at import
time is what turns a missing .git into a collection error, which is the
failure mode being removed here, so it is not the form to copy.

So the kit_repo_only marker cannot help this tree at all: the exception fires
at import, long before any marker is consulted. #203 owns consolidating those
three copies, and doing so is a prerequisite for a no-.git tree collecting.

PR #232 removed one contributor to this (a module-scope
config/dev-model.yaml read in test_init_sh.py, now lazy) but deliberately
does not claim to fix the no-.git case, and its measured trees all have
.git — which every real adopter repository does by definition. This is the
tarball/archive path.

Related: #203, #134, #232.