…arpdotdev#9870)

On Linux, agents distributed as Node.js or Python shebang scripts surface
as `node /usr/local/bin/codex` (the runtime, not the script) because
`/proc/PID/comm` is initialised from the executable basename. The bare
prefix match in `CLIAgent::detect` saw `node` and missed Codex entirely,
so the third-party CLI agent toolbar never appeared.

The reproduction in warpdotdev#9870 confirms this is the canonical case: the
@openai/codex npm package installs `/usr/local/bin/codex` as a symlink to
`codex.js` whose shebang is `#!/usr/bin/env node`. Detection failed for
every Linux user of Codex, but worked on macOS where shebang process
naming differs.

Fix:

- New helper `path_basename_token` peels a leading path component so
  `/usr/local/bin/codex` → `codex` for matching purposes.
- New helper `shebang_script_basename` recognises a known script-runtime
  invocation (`node`, `nodejs`, `bun`, `deno`, `python`, `python3`) and
  pulls the script basename from `argv[1]`, stripping common script
  extensions (`.js`, `.mjs`, `.cjs`, `.ts`, `.py`).
- `detect()` matches `command_prefix()` against the exact first word OR
  either of the two derived tokens.

The change is purely additive — every existing detect path still matches
exactly as before, including the Uber-team `aifx agent run claude`
special case. The new candidates only fire when the exact-word match
fails, so there's no false-positive surface for unrelated commands like
`node ./my-app.js` or `python /home/user/script.py`.

Tests in `cli_agent_tests.rs` cover:

- `node /usr/local/bin/codex` → Codex (the canonical warpdotdev#9870 case)
- `node` with flags before the script path
- `node /path/to/codex.js` (extension stripped)
- `nodejs`, `bun` runtimes
- Path-prefixed binaries: `/usr/local/bin/codex`, `./goose`
- Negative tests: `node`, `node ./my-app.js`, `python script.py`,
  `node --version` all return None.

Followups deliberately out of scope:
- A user-configurable runtime list (V0 hardcodes the 6 most common).
- An `argv[0]` rewrite via `process.title` patches upstream — that's a
  Codex CLI repo concern, not a Warp concern.

Resolves the merge conflict with master after warpdotdev#9667 (Mistral Vibe agent
recognition) merged. Combines both surfaces:

- The Vibe arm now also resolves through path_basename_token and
  shebang_script_basename, so `/usr/local/bin/vibe-acp` and
  `node /opt/vibe-acp` both detect as Vibe.

Also addresses oz-for-oss review on PR warpdotdev#10022:

> This treats the first non-flag token as a script even when the
> preceding runtime flag consumes an argument, so invocations like
> `node -e codex`, `node --require codex`, or `python -c claude` can
> false-positive as agents.

Adds VALUE_TAKING_RUNTIME_FLAGS, a list of flags that consume the
following token as their value:
- Node: -e, --eval, -p, --print, -r, --require, -C, --conditions
- Python: -c, -m, -X, -W
- Deno (subset): -A, --allow-read, --allow-write, --allow-net
- Bun (subset): -d, --define

shebang_script_basename now consumes the value alongside the flag, so
`node -e codex` no longer false-positives as Codex. The `--key=value`
form remains a single whitespace token and is handled by the existing
`starts_with('-')` flag check.

New regression tests in cli_agent_tests.rs:
- `node -e codex` → None
- `node --eval codex` → None
- `node --require claude /home/user/app.js` → None
- `python -c claude` → None
- `python -m gemini` → None
- `node --require some-mod /usr/local/bin/codex` → Codex (legitimate
  agent script after value-taking flag is consumed)
- `node --inspect=127.0.0.1:9229 /usr/local/bin/codex` → Codex
  (--key=value form preserves correct script detection)

Addresses two findings from oz-for-oss re-review on PR warpdotdev#10022:

1. `CLIAgent::CursorCli`'s `command_prefix()` is the bare string "agent",
   which is generic enough to false-positive when the basename / runtime-
   script fallbacks fire on unrelated scripts and binaries. `/tmp/agent`,
   `node /tmp/agent.js`, `python /tmp/agent.py` would all incorrectly
   detect as Cursor.

   New `GENERIC_AGENT_PREFIXES` const lists prefixes that are too generic
   for fallback recovery; `prefix_eligible_for_basename_recovery` gates
   the candidate_basename and runtime_invoked_basename branches in
   `CLIAgent::detect`. Exact-word match still works for Cursor, just not
   the basename fallbacks. Distinctive prefixes (codex, claude, etc.)
   continue to recover via basename normally.

2. `shebang_script_basename` did its own whitespace tokenization, which
   diverged from `extract_first_command`'s shell-aware parsing when the
   user's command starts with env-var assignments. `FOO=1 node /usr/local/
   bin/codex` would have `resolved_first_word == "node"` (correct) but
   `shebang_script_basename` would read `FOO=1` as the runtime token and
   never reach the script.

   Added `is_env_var_assignment` helper (matches POSIX env-var name
   pattern: `[A-Za-z_][A-Za-z0-9_]*=`). `shebang_script_basename` now
   skips leading env-var assignments before reading the runtime token,
   keeping it aligned with `extract_first_command`.

New regression tests:
- `test_detect_generic_prefix_no_basename_false_positive`:
  - `agent` → CursorCli (exact match still works)
  - `/tmp/agent`, `node /tmp/agent.js`, `python /tmp/agent.py` → None
  - `/tmp/codex` → Codex (distinctive prefixes still recover)
- `test_detect_node_shebang_with_env_var_prefix`:
  - `FOO=1 node /usr/local/bin/codex` → Codex
  - `FOO=1 BAR=baz node /usr/local/bin/codex` → Codex
  - `PATH=/tmp:/usr/bin node /usr/local/bin/codex` → Codex