fix(beads): use Bash escapes for recovery command controls (FUZZ-040 r2)

[randlee]

Fixes comp-t-7.fuzz-040-r2-fix (round 2 of FUZZ-040, filed by comp-t-7.15-qa): recovery-command shell quoting uses bash $'...' with fixed-width \xHH / \uHHHH / \UHHHHHHHH escapes for control and bidi format characters (U+202A-202E, U+2066-2069, U+200E/F), never Rust escape_default. Regression tests cover the library and CLI paths with a bash round-trip.

Rebased 5e99ce0 -> fe1e05b onto #718. Gates at fe1e05b: clippy -D warnings clean, fmt clean, cargo test --workspace 1067 passed / 0 failed / 0 ignored.

🤖 Generated with Claude Code

[randlee]

QA fix verification: comp-t-7.fuzz-040-r2-qa (FUZZ-040, round 2) at fe1e05b

Verdict: PASS on the carried finding's own criteria. Two follow-up observations are reported, not held against the fix.

Finding	Disposition
fuzz-040	fixed (see judgement note)

Verified by the fuzz coordinator at the pinned commit:

• ASCII control characters (BEL, ESC, newline, tab, DEL, 0x01, 0x1F) print as $'\xHH'; U+0085 and the bidi characters U+202A-202E, U+2066-2069, U+200E/F print as $'\uHHHH'. No raw bidi or control character appears in the output, and all of them round-trip byte for byte under bash 5.3 (bash -c 'printf %s <arg>'); the control characters also under macOS /bin/bash 3.2.
• Single quote, backslash, $(id)58, semicolon, spaces, an accented letter, an emoji and literal \x41 / ‮ text all round-trip. A plain id prints as 'spc-1', the round-1 form.
• The regression test now drives BeadComposeError::GraphEdgeMissing and missing_edge_recovery_commands with a hostile id (control character and U+202E), asserts no raw control or listed bidi character, and does a dynamic bash round-trip. It is un-ignored, passes, and fails at the round-1 commit 1f37cde (old \u{7} form). cargo test -p sc-compose passes.
• Local fmt, clippy -D warnings and tests are clean at this head.

Judgement note. The coordinator returned open. I closed it fixed because its two remaining defects lie outside the finding's own wording ("every printed argument shell-quoted, control characters escaped", bidi named in the observation); they came from characters I added to the verifier's check list, not from the carried finding. Team-lead can reopen. Observations, for a follow-up bead if wanted:

• fuzz-040-r2-a: needs_shell_escape covers is_control plus a hand-picked bidi list, so other format and line-separator characters (U+2028, U+2029, U+FEFF, U+200B, U+200D, U+2060, U+00AD, U+061C, U+E0001) are still printed raw inside the quotes. Round-trip is exact, but they are raw in the terminal. Suggest escaping all Unicode Cf plus U+2028/2029.
• fuzz-040-r2-b: \uHHHH is not valid in macOS /bin/bash 3.2, so the escaped bidi characters and U+0085 print as literal text there (zsh and bash 4.2+ are fine). Encoding as \xHH UTF-8 bytes would work everywhere. The test runs whichever bash is first on PATH.

Disclosures: no real bd. The CLI human text path was checked through the bead.rs unit test, not by running bead preview-attach or attach, and --json receipt output was not compared byte for byte (the code gates the new diagnostics on !json, and shell_quote is not on the JSON path). The human refusal message line of the FUZZ-039 layer below, which can print raw control/bidi characters from an id, was not examined here. CI not checked (rate-limit rule). Hand-locked: fix-round-scope accepted the coordinator for fuzz-040.