refactor(demo-26): S14 PR 1 — failure modes on resilience libraries

[patterncatalyst]

S14 PR 1 of 4 (plan: _plans/s14-resilience-plan.md, steps 0 and 1). Example 26's breaker, retry and bulkhead move from hand-written code to each stack's resilience library.

Stack	Breaker	Retry	Bulkhead
Spring Boot	Resilience4j 2.4.0 @CircuitBreaker + fallback method	@Retry	@Bulkhead
Quarkus	SmallRye Fault Tolerance @CircuitBreaker + @Fallback	@Retry + @ExponentialBackoff	@Bulkhead
.NET	Polly v8 pipelines (Microsoft.Extensions.Http.Resilience 10.10.0)	HttpRetryStrategyOptions	concurrency limiter, no queue
Python	pybreaker 1.4.1 through calling()	tenacity 9.2.1 AsyncRetrying	asyncio.Semaphore
Go	gobreaker v2 (unchanged; trial calls now 2)	cenkalti/backoff v7 (unchanged)	x/sync/semaphore TryAcquire
C++	hand-written (no maintained library on ConanCenter, D3)	coroutine loop with sleepCoro	std::counting_semaphore

Contract changes (user decisions D1, D2):

• D1: a full bulkhead rejects at once with 429 in every language. Before, Spring, Quarkus, .NET and Python answered 200 with an error body, and Go and C++ answered 503. verify.sh adds a saturation check and a recovery check, taking 26 from 20 to 22 checks.
• D2: /breaker-state is {name, state} plus whatever metrics the library exposes; the state is closed, open or half-open.
• The shared breaker settings are 5 failures to open, 10 s open, and 2 trial successes to close. Polly lets 1 trial call through by design.
• Quarkus and Go backends move to /process with X-Deadline-Ms, like the other four.

Clean-ups found on the way:

• The C++ edge used blocking libcurl calls and sleep_for on Drogon event-loop threads. It now uses coroutine handlers and Drogon's async HttpClient, so libcurl is gone (lock cleaned).
• The C++ backend's slow mode now answers from a timer.
• Python opens one httpx.AsyncClient in the lifespan.
• .NET pins Polly 8.8.0 directly, because Microsoft.Extensions.Http.Resilience only sets a >= 8.4.2 floor.

Docs:

• Appendix M's timeout, retry and breaker codetabs and the 26 READMEs are now condensed excerpts of the examples.
• In all six decks, the 18 Appendix M code slides became excerpts with matching chips and notes. They previously showed illustrative code that no example backed. All 18 were rendered and read, and deck_fit reports 0.

Verification:

• Baseline on main was 36/36 (examples 02, 03, 11, 12, 18, 26).
• verify-matrix.sh 26-failure-modes on the final commit: 6/6, 22/22 each.
• validate_site, voice-scan and deck_text --check report clean.