test(e2e): verify the herdr flow end to end + fix nix run .#e2e on macOS

[jakemassoth]

What this is

A verification/QA pass proving the new herdr session/TUI flow (PR #29) works end to end, plus one trivial fix uncovered along the way.

All testing ran in a fully isolated sandbox — a unique herdr session (yt-e2e-9021), a private XDG_CONFIG_HOME and YESCHEF_HOME under /tmp, never the live yeschef brigade or the shared config home. The live brigade was confirmed running and untouched before and after.

The one fix (this branch's only diff)

nix run .#e2e — the documented e2e command (DEVELOPMENT.md) and a CI step — failed on macOS before running a single test:

error: linking with `cc` failed: ... note: error: tool 'clang' not found
error: could not compile `serde_core` (build script) ...

Root cause: the e2e app's runtimeInputs shipped only rustToolchain + herdr + git. The pinned rust toolchain needs a C compiler/linker to build proc-macro build scripts and link the test binary; on macOS there is no ambient nix cc-wrapper on PATH, so the link fails. (Linux CI has a system cc, which masks it.) Fix: add pkgs.stdenv.cc — the same clang-wrapper the devShell already pulls in.

• Before: nix run .#e2e → tool 'clang' not found, 0 tests run.
• After: nix run .#e2e → compiles, 19/19 tests pass.
• nix flake check → fmt / nixfmt / lint / unit tests all ✅.

What was verified (end to end)

Automated: the full e2e suite passes 19/19 (via nix run .#e2e after the fix, and via the devShell).

Manual walkthrough against a throwaway git repo + a dummy --agent shell loop (no real coding agent, no external services), with captured evidence for every stage:

Stage	Result
init	home + yeschef.db + AGENTS.md created
project add / list	bare clone made, origin/main resolves, project listed
spawn	git worktree created & tracking origin/main; prompt file written outside the worktree (status preamble + verbatim brief); herdr workspace demo/widget created; agent launched via the short "read the brief" indirection
peek	pane read returned the marker + the file-indirection launch instruction
send	one-line steer reached the pane (echoed back by the cook)
status	ticket table joins the registry with herdr's live agent_status + self-reported task status
ticket status-set	task status persisted (NEW → IN_PROGRESS → DONE) and reflected in status
restart	herdr server bounced; both workspaces restored from herdr's persisted session state (headchef + demo/widget), server back up on the isolated socket
kill --rm-worktree	workspace closed, worktree removed, ticket deregistered
cleanup (dry-run)	correctly flagged "merged and status DONE", removed nothing
cleanup --yes	reaped the ticket (workspace closed, worktree + branch removed, deregistered)

What does NOT work / notes

• tui / attach are interactive passthroughs to herdr's native UI (they hand over the terminal / block on a TTY), so they were verified by code inspection + the fact that the server they attach to is proven working — not exercised headlessly.
• restart restores workspace shape for our dummy agent, not the process (expected — our sleep/cat stand-in isn't a herdr-recognized agent, so resume_agents_on_restore doesn't apply; real Claude Code cooks resume natively).
• herdr reports our dummy agents as STATE = unknown (expected — sh/cat aren't recognized coding agents).

No other defects found. The herdr flow works as designed.