build(codegen): optional STR_KILL_PRE fallback to unlock 2.0.x codegen (deferred)

[CameronBrooks11]

Deferred — do NOT build now

Lane B verdict (design doc): a single committed _constants.py from one pinned
2.1.x SHA is sufficient. All 12 tracked message literals are byte-identical
across 2.0.9.3 → 2.1 → 2.1.1 → 2.1.2 → latest-2.1.2.x → bugfix-2.1.x. Exactly one
tracked macro drifts — STR_KILL_PRE — and only at the 2.0.x↔2.1.x boundary, where
it was introduced (not renamed). gen_constants.py:79-81 raises SystemExit on a
missing wanted macro, so pinning codegen to a 2.0.x ref fails loudly rather than
producing a wrong table. Runtime already degrades gracefully on 2.0.x: the 2.1.x
marker "!! KILL caused by " is a substring of the 2.0.9.3 inlined kill messages,
so substring-based FATAL_MARKERS detection (protocol.py:124) still fires.

Proposed approach (only if 2.0.x codegen is ever actually wanted)

Do not build a per-version _constants/<version>.py matrix — unwarranted for one
optional macro across a major boundary. The cheapest fix is a one-line fallback in
scripts/gen_constants.py: treat STR_KILL_PRE as optional, defaulting to
"!! KILL caused by " when absent. That lets codegen target either line from one
table.

Acceptance criteria (when/if taken)

• STR_KILL_PRE becomes an optional wanted macro with a default; codegen from a
  2.0.9.3 ref succeeds and produces a correct table.
• No matrix module is introduced.

Evidence

scripts/gen_constants.py:26-39,57-66,79-81, marlin_host/_constants.py,
marlin_host/protocol.py:124.
Design: working/dialect-and-mock-investigation.md (Later).

Part of #5.