refactor(sql_builder): consolidate slot SQL preparation into dedicated functions

[isdaniel]

• Deserializer error policy (deserializer.rs): collapse the two ~220-line
  serde impls behind a compile-time ErrorPolicy trait (Strict/Lenient) plus one
  shared coercion impl. Coercion rules and error messages are now single-sourced.
  Strict's zero-copy visit_borrowed_* + #[inline] path is preserved via
  monomorphization. The two Deserializer impls stay separate (a single impl is
  impossible under serde's 'de borrow model) — they differ only in the final
  borrowed-vs-owned visit.
• Slot glue (sql_builder.rs, both backends): add
  prepare_{create,alter,read}_slot to pair the version-check with the SQL builder
  in one place; both backends now call them instead of re-pairing. The real seam
• Frame-injection test seam (libpq.rs, stream.rs): add libpq's matching
  null_for_testing_with_frames; 8 shared-pump tests now run on both backends,
  closing the libpq COPY-ingress coverage hole. ..._terminates_after_transport_stop
  stays rustls-only (it drives PQputCopyEnd, which segfaults on libpq's
  null-pointer conn; that path is backend-specific and integration-tested).

[isdaniel]

@copilot review

[codspeed-hq]

Merging this PR will not alter performance

✅ 50 untouched benchmarks

---

_{Comparing refactor/collapse-deserializer-policy-slot-glue-test-seam (235752d) with main (df60b02)}

[codecov]

Codecov Report

❌ Patch coverage is 98.02632% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 94.96%. Comparing base (df60b02) to head (235752d).

Files with missing lines	Patch %	Lines
src/deserializer.rs	98.52%	4 Missing ⚠️
src/sql_builder.rs	91.30%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #109      +/-   ##
==========================================
+ Coverage   94.76%   94.96%   +0.20%     
==========================================
  Files          27       27              
  Lines       20678    20805     +127     
==========================================
+ Hits        19596    19758     +162     
+ Misses       1082     1047      -35     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

@copilot review

Reviewed 5805412; no actionable correctness or security findings. Local fmt, both backend unit-test suites, and Clippy pass. The Fuzz check failure is a nightly rustc ICE while compiling Tokio, not a project-code failure.

[isdaniel]

@copilot review

[Copilot]

@copilot review

Second review pass done. Verified in more depth this time:

• Lenient numeric semantics preserved: lenient now routes through parse_int_signed/parse_int_unsigned instead of FromStr. Checked the edge case — parse_int_signed accumulates negatives via checked_sub, so i64::MIN still parses (covered at deserializer.rs:2974). Invalid-UTF-8 integer input still yields the invalid UTF-8 for {type} message, since numeric_parse_msg validates on the cold path.
• Lenient bytes/byte_buf/unit/enum are behavior-identical to the pre-refactor impls (same messages, same fallbacks, single error push).
• Slot glue: prepare_* keeps check-before-build ordering in both backends.
• libpq test seam is inside a #[cfg(test)] impl, so it isn't dead code in release builds.

One note: lenient-mode message wording changed for two cases — integer parse failures no longer append the FromStr cause, and invalid bool now reports the (expected t/f/...) form via lossy_token. That's the intended single-sourcing with strict, and no test asserts the old text.

Fixed two comment defects introduced when the commit was amended/reflowed: a duplicated sentence fragment in stream.rs and a garbled sentence (missing word) plus the dropped doc line on prepare_create_slot in sql_builder.rs — dd7f4c2. fmt, clippy, and both backend suites (1246 libpq / 1240 rustls) pass.

Fuzz CI remains red from a nightly rustc ICE while compiling Tokio, unrelated to this branch.

@copilot review

Correction: the comment fixes landed in 8cbab4a, not the hash cited above.

[isdaniel]

@copilot review

[Copilot]

@copilot review

Reviewed 235752d; no actionable correctness or security findings. Local fmt, both backend unit-test suites, and Clippy pass. CI and Benchmarks pass, and the pinned-nightly Fuzz workflow now passes all three targets; Code Quality is still running.