test: pin the branch in the t900 tests that push to main

[mizchi]

Fixes #193, #194, #195.

Three tests create their repositories with a bare git init and then name main, so they fail wherever init.defaultBranch is unset and git still defaults to master (git 2.43 here).

	before	after	failing assertion
t9001-subdir-push	19 of 23	23 of 23	git push origin main (×3, plus one knock-on)
t9004-workspace-routing	7 of 8	8 of 8	grep "main" branch.out
t9005-workspace-commit-push	6 of 7	7 of 7	git push origin main

The fix

For t9001 and t9005 the branch to pin is the bare upstream's, not the clone's — the working clone adopts the remote's unborn HEAD, so one change fixes the push without touching the clone:

 mkdir upstream.git &&
-(cd upstream.git && git init --bare) &&
+(cd upstream.git &&
+ git init --bare &&
+ git symbolic-ref HEAD refs/heads/main) &&
 git clone upstream.git work &&

t9004 creates its repository directly, so it is pinned in place.

Setting HEAD while it is unborn rather than passing git init -b, which needs git 2.28 — the same form used for the fixture in #192.

Verification

All three pass in full with init.defaultBranch unset, and with it set to main — so the fix removes the dependency rather than moving it:

                              unset       =main
t9001-subdir-push             23 / 0      23 / 0
t9004-workspace-routing        8 / 0       8 / 0
t9005-workspace-commit-push    7 / 0       7 / 0

The whole test-subdir task afterwards:

t900:  Total 11   Passed 10   Failed 1
t902:  Total  2   Passed  2   Failed 0

The one remaining failure is t9007-workspace-nested-translation.sh, left alone deliberately — see #196. It expects bit repo status from a deep nested path to fail with Not a git repository and it does not, while the sibling assertion at the workspace root passes. Deciding whether the test's expectation or the behaviour is right is a call for a maintainer, not a mechanical fix.

A correction worth flagging

While preparing this I found that four of the six issues I filed after #192 were wrong, and I have corrected them: #197 and #198 are closed as not-a-bug, and #194, #195, #196 had retractions added.

I had reported that implicit command translation was broken — bit status inside a workspace returning plain git status instead of the workspace form. It is not. The container I triaged in exports git config through the environment (GIT_CONFIG_COUNT, GIT_CONFIG_KEY_0..2, GIT_CONFIG_VALUE_0..2, holding credential.interactive and two url.*.insteadOf entries for its proxy). should_delegate_to_real_git (modules/bit/cmd/bit/main.mbt:1044) treats any such injection as a reason to hand the command to the real git binary, and it runs before workspace_translate_implicit_command — which is the documented intent of that check. With those variables unset, every one of those assertions passes.

Every measurement in this PR was taken with GIT_CONFIG_* unset.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VDVUJHu38YtVAr6sKZtRDB

---

Generated by Claude Code