… exit 1

Replaces both `exit 1` paths in the Playwright install/verify block with a
$_PW_FAIL_REASON accumulator that prints a named warning to stderr and
returns control. Previously a Chromium download failure (network blip,
locked-down corporate machine, missing Node.js on Windows) would abort
./setup mid-run with no skills registered, no hooks installed, no
migrations applied — and a re-run would start over from the top.

Now the failure only disables browser-driven features (/qa,
/design-review, make-pdf, sidebar, /pair-agent); skills, hooks, and
migrations land normally. The warning names which sub-step failed
(chromium-install / windows-no-node / windows-node-modules /
post-install-launch) so users and bug reports can be specific, and
tells the user to re-run ./setup to retry just this step.

Tests (5) exercise the actual block from setup with stubbed
ensure_playwright_browser and bunx, simulate install + post-install
failures, and assert the script exits 0 with the named warning. One
side-by-side test inlines the OLD bug shape and confirms it returned
non-zero, so the exit-code difference is proven, not just asserted.

Coordinates with garrytan#1838 (pins Playwright install version): both touch the
same code block, so whichever lands first needs the other to rebase. The
changes are complementary — garrytan#1838 fixes which Chromium build gets
installed; this PR fixes what happens when the install fails.

Carved out from garrytan#1883 per @jbetala7's review request.

…ability

setup used to interleave trivial filesystem ops, expensive compute,
settings.json mutations, and network/sudo steps in whatever order the
script grew. A Playwright Chromium download failing at step 2 would
abort before skills got registered at step 4, hooks landed at step 10,
or migrations ran at step 8.

This commit groups work by where failures actually come from:

  Phase A — filesystem (always succeeds): ~/.gstack/, welcome, GBrain detection
  Phase B — local compute (low fail): bun build, skill linking, migrations, regen
  Phase C — settings.json mutation (medium fail, reversible): team-mode + plan-tune hooks
  Phase D — external network/sudo/package managers (highest fail): coreutils, font, Playwright

Each phase is bracketed with a header comment; steps within are
renumbered (A1, A2, B1, B2a...).

The biggest behavior shift is the GBrain detect/regen split that the
new order requires: detection moves to A3 (filesystem scan only, no
dependencies), regen moves to B6 (runs after the binary is built and
skill docs are generated). Detection persists to
~/.gstack/gbrain-detection.json so B6 can read it.

A safety guard for that split: when gstack-gbrain-detect exits non-zero,
A3 removes both the temp file AND any pre-existing detection file from a
prior successful run. Without the guard, a stale "ok" detection file
would survive a failed detect and silently apply ~250 tokens of overhead
per planning skill to a user whose gbrain may no longer be installed.

Behavior preservation:
- bun_cmd routing in link_{codex,factory,opencode}_skill_dirs is
  unchanged (still buggy here). The fix is in garrytan#1898.
- gen:skill-docs:user still pipes through `tail -3` in B6 (still buggy
  here). The fix is in garrytan#1898.
- Playwright Chromium install/verify in D3 still exits 1 on failure
  (unchanged). The warn-don't-exit conversion is in garrytan#1900 and
  coordinates with garrytan#1838.

Splitting those keeps this PR reviewable as "a move plus the gbrain
split that the move requires," per @jbetala7's request on garrytan#1883.

… exit 1

Replaces both `exit 1` paths in the Playwright install/verify block with a
$_PW_FAIL_REASON accumulator that prints a named warning to stderr and
returns control. Previously a Chromium download failure (network blip,
locked-down corporate machine, missing Node.js on Windows) would abort
./setup mid-run with no skills registered, no hooks installed, no
migrations applied — and a re-run would start over from the top.

Now the failure only disables browser-driven features (/qa,
/design-review, make-pdf, sidebar, /pair-agent); skills, hooks, and
migrations land normally. The warning names which sub-step failed
(chromium-install / windows-no-node / windows-node-modules /
post-install-launch) so users and bug reports can be specific, and
tells the user to re-run ./setup to retry just this step.

Tests (5) exercise the actual block from setup with stubbed
ensure_playwright_browser and bunx, simulate install + post-install
failures, and assert the script exits 0 with the named warning. One
side-by-side test inlines the OLD bug shape and confirms it returned
non-zero, so the exit-code difference is proven, not just asserted.

Coordinates with garrytan#1838 (pins Playwright install version): both touch the
same code block, so whichever lands first needs the other to rebase. The
changes are complementary — garrytan#1838 fixes which Chromium build gets
installed; this PR fixes what happens when the install fails.

Carved out from garrytan#1883 per @jbetala7's review request.

… exit 1

Replaces both `exit 1` paths in the Playwright install/verify block with a
$_PW_FAIL_REASON accumulator that prints a named warning to stderr and
returns control. Previously a Chromium download failure (network blip,
locked-down corporate machine, missing Node.js on Windows) would abort
./setup mid-run with no skills registered, no hooks installed, no
migrations applied — and a re-run would start over from the top.

Now the failure only disables browser-driven features (/qa,
/design-review, make-pdf, sidebar, /pair-agent); skills, hooks, and
migrations land normally. The warning names which sub-step failed
(chromium-install / windows-no-node / windows-node-modules /
post-install-launch) so users and bug reports can be specific, and
tells the user to re-run ./setup to retry just this step.

Tests (5) exercise the actual block from setup with stubbed
ensure_playwright_browser and bunx, simulate install + post-install
failures, and assert the script exits 0 with the named warning. One
side-by-side test inlines the OLD bug shape and confirms it returned
non-zero, so the exit-code difference is proven, not just asserted.

Coordinates with garrytan#1838 (pins Playwright install version): both touch the
same code block, so whichever lands first needs the other to rebase. The
changes are complementary — garrytan#1838 fixes which Chromium build gets
installed; this PR fixes what happens when the install fails.

Carved out from garrytan#1883 per @jbetala7's review request.