A caller can make a write conditional on the state of the block it read

[muellerei]

Closes #123

Summary

get-block --json                      -> hash, tree_hash
find-block / get-page / get-todos /
get-journal-range  --json             -> hash on every block

update-block, set-todo-status, set-block-property, remove-property --id,
move-block, copy-block --remove, add-journal-block --upsert-heading
  --expect-hash H          remove-block / delete-block  --expect-tree-hash H
  set-todo-status          --expect-marker M

WriteCommand(owes=…)           the one place that asks "did the call say what it read?"
  guard_write()                read_only first
  guard_preconditions()        [safety] require_preconditions
  command body
    read past the cache
    check_precondition()       before --dry-run, the no-change answer and the first write
    update_block(expect=…)     the write itself refuses with block_changed if it moved again

 [safety]
 read_only = false
+require_preconditions = true     # also --require-preconditions, LOGSEQ_CLI_REQUIRE_PRECONDITIONS

Evidence

• Before: an agent reads a block, the user edits it in Logseq, the agent's write goes through over the edit. Suite: 3793 tests.
  After: the same call refuses with reason: precondition_failed (shown from the test double), nothing written:

$ logseq-cli update-block --id 6650a1b2-0000-4000-8000-000000000001 \
    --content "TODO ship the parser, fixed" --expect-hash 3fa9c01b27de
Error: Block 6650a1b2.. changed since it was read (marker is TODO, first line:
"TODO ship the parser"). Read it again and decide again. Nothing was written.
$ echo $?
1

• The hash ignores collapsed:: and CLOCK lines, so folding a block or a running clock does not fire it.
• Suite: 4029 passed, 8 skipped, on Python 3.10, 3.11, 3.12 and 3.13; ruff check . clean; the wheel check passes.
• Each new rule has a test that fails when the rule is removed (the check, the cache bypass, the owes mark, the always text, the order against read_only, the Upsert exception).
• The hash adds about 30 bytes per block to --json reads.

Merge Danger

Door: two-way. The switch is off by default and the new options are optional; reverting the commit removes them.

Blast Radius: JSON output

hash (and tree_hash on get-block) are new fields in the JSON of five read commands, and --max-chars counts them. Scripts that compare whole JSON blocks will see the new key. Four calls that cannot use a precondition now say so (--expect-hash without --upsert-heading, with several --content, on remove-property --page, on copy-block without --remove) with a usage error instead of ignoring the option.

Known limits, stated in docs/safety.md: no compare-and-swap, the hash sees content and not position, write outputs carry no new hash yet, and copy-block --remove does not cover the window in which it copies.