fix: three filed bugs where codexclaw obstructed its own instructions (#47, #48, #49)

[lidge-jun]

Promotes dev to main for 0.2.12: three filed bugs, all of them cases where
codexclaw obstructed an agent that was following its own instructions.

#47 — the sibling commands had no --help

cxc --help points at loop, scan and receipt. Following that pointer
failed: --help was an unknown verb on all three, and cxc --version was an
unknown command. orchestrate was fixed for exactly this in
260709_cxc_help_agent_ux; its siblings never were.

The individual error messages were fine. The defect is that discovery was only
available through failure — arming a goalplan in this session took six
consecutive rejections to assemble one command:

loop steer: --session <id> is required
loop steer: --batch-json <path-or-json> is required
loop steer: idempotencyKey is required and must be a non-empty string
loop steer: rationale is required and must be a non-empty string
loop steer: evidence is required and must be a non-empty string
loop steer: ops must be a non-empty array

The loop usage now spells out the steer batch shape. scan record also accepts
--cwd, which orchestrate already documented.

Worth noting: cxc --version worked from one entry point and still failed from
the other, because there are two. The test asserts both — that is the mistake it
exists to catch.

#48 — one session id, two FSMs

Session files live at <cwd>/.codexclaw/sessions/<id>.json. The id is stable,
the cwd is not, so a thread whose cwd is one tree while its work is in another
has two FSMs under one id. The reporter closed D in the wiki tree and the next
turn re-injected Interview from the other copy.

Pinning the store to the session-start workspace would invalidate every existing
session file with no migration path, so this makes the split visible instead:

session=<id> phase=IDLE ...
WARNING: this session id also has state in 1 other tree(s); the phase above
describes THIS cwd only.
  also at: .../cxc-split-a/.codexclaw/sessions/<id>.json

Detection only — the other tree is never read from or written to. loop show
also accepts --session and resolves the slug the session already carries.

#49 — two gates that forced the forgery they prevent

receipt test refused a receipt whenever the check dirtied the tree, including
when the dirty files were the artifacts the check exists to rebuild. The reported
workaround was to commit the generated files and run a no-op existence check — a
receipt that certifies nothing. A forged receipt is strictly worse than a loose
one, because it satisfies CHECK-BINDING-01 while proving less than no receipt.

--generated <path> declares expected rewrites. Everything undeclared is still
refused, verified four ways including that a declared FILE does not cover its
siblings.

orchestrate D refused a goalplan whose work-phases were all done, because
"complete" and "empty" produced the same internal result. The workaround was to
write a finished phase back to in_progress purely to pass the gate — corrupting
the record to satisfy a check about the record.

SOURCE-DELTA-01 was raised in the same issue and deliberately left alone: I hit
it twice while doing this work and both times it was right.

Verification

• native Windows: 1961 tests, 1953 pass, 0 fail, 8 skip
• WSL Ubuntu: 1961 tests, 1960 pass, 0 fail, 1 skip
• --generated verified end to end on a throwaway git repo whose validator
  rewrites its own output: refused without the flag, receipt written with it,
  still refused when the wrong path is declared
• the #48 warning verified against two real trees holding the same id