feat(snapshot): add --snapshot-update to re-record snapshots

[Chemaclass]

Context

There is no way to re-record snapshots. The documented procedure is to delete the
snapshot file and re-run, since assert_match_snapshot writes the file when it is
missing (src/assert_snapshot.sh:24-27). Both agent-facing docs say so
explicitly:

• docs/ai-agents.md:119 — "There is no --update-snapshots flag"
• docs/public/bashunit-skill.md:175 — same

The path to delete is derived, not stated in the test source
(bashunit::snapshot::resolve_file, src/assert_snapshot.sh:91-116): the test
file's basename and the test function name are each passed through
normalize_variable_name, then joined as
./<dir>/snapshots/<file>.<function>.snapshot. For example:

snapshots/snap_test_sh.test_new_snapshot_is_silently_recorded.snapshot

So re-recording means reproducing that normalization by hand, per test.

Proposal

Add --snapshot-update (plus BASHUNIT_SNAPSHOT_UPDATE for .env, following the
existing flag/env pairing in src/env.sh) that makes assert_match_snapshot and
assert_match_snapshot_ignore_colors overwrite the snapshot with the actual value
and count it as a recorded snapshot rather than a pass.

Details worth settling:

• it must compose with a filter so a single test can be re-recorded, not the whole
  suite — --snapshot-update together with the existing --filter
• interaction with placeholders (::ignore::, BASHUNIT_SNAPSHOT_PLACEHOLDER):
  overwriting a snapshot destroys its placeholders. Either refuse to overwrite
  a snapshot containing the placeholder, or warn loudly. This needs a decision
  before implementation.
• run summary must make the rewrite visible (reuse the existing
  Some snapshots created block, wording adjusted)

Why this matters for agentic coding

Deliberate output changes are routine — an agent adjusts a message and a dozen
acceptance snapshots need re-recording. Today the agent must derive each path
through a normalization function it has to read the source to understand, then
rm files it cannot verify it chose correctly. Deletion is also the one operation
with no undo: .claude/skills/gh-issues/SKILL.md and the skill docs both already
warn that a deleted snapshot never fails, so a wrong rm silently converts a real
assertion into a rubber stamp. A flag makes the intended operation the easy one
and takes rm off the agent's path entirely.

Acceptance criteria

• --snapshot-update rewrites existing snapshots from the actual value
• BASHUNIT_SNAPSHOT_UPDATE equivalent, documented in docs/configuration.md
• Composes with --filter for single-test re-recording
• Placeholder-bearing snapshots are protected or loudly warned about
• Rewritten snapshots are reported in the summary, not counted as passes
• Docs updated: docs/snapshots.md, docs/command-line.md,
  docs/configuration.md, and the "there is no such flag" notes in
  docs/ai-agents.md + docs/public/bashunit-skill.md
• Acceptance tests in tests/acceptance/
• Bash 3.0+ compatible, make sa + make lint green

[Chemaclass]

Done in #909 (merged).