[AZ-505] Cycle 6 Step 17: retrospective + close cycle

Single-task cycle delivering AZ-505 (3 SP); 1 batch, PASS verdict
after a single auto-fix round (ComputeLocationHash duplication
consolidated into Uuidv5.LocationHashForTile). Step 14 Security
Audit skipped; Step 15 Performance Test PASS (8/8, exit 0) and
closes the cycle-3 perf-harness leftover that carried across
cycles 3-5.

Top 3 lessons appended to LESSONS.md ring buffer:
- Kestrel Http1AndHttp2 requires TLS for ALPN; spec-time decision
- Dapper-bypassing test paths own the Npgsql type contract
- Test fixtures naming specific schema artifacts need migration awareness

Cycle 7 opens at Step 9 (New Task).

Co-authored-by: Cursor <cursoragent@cursor.com>

_docs/LESSONS.md

@@ -37,6 +37,12 @@ If the enum's wire string happens to match a member name case-insensitively (e.g
 
 ## Ring buffer (last 15 entries — newest at top)
 
+- [2026-05-12] [tooling] Kestrel `HttpProtocols.Http1AndHttp2` silently serves only HTTP/1.1 over a plaintext listener — ALPN requires TLS, so any "enable HTTP/2" task without TLS in its definition-of-done will downgrade transparently and the only log line is at INFO; tasks that mention HTTP/2 / h2 / multiplexing / ALPN MUST resolve the TLS-vs-h2c choice at spec-write time and the test gate MUST assert `HttpVersion == 2.0` not just a 200 (cycle 6: AZ-505 AC-5 first landed on h2c plaintext, required a post-merge TLS+ALPN pivot with dev-cert plumbing across compose/tests/perf/docs).
+  Source: _docs/06_metrics/retro_2026-05-12_cycle6.md
+- [2026-05-12] [testing] When a test bypasses Dapper to gain access to a feature Dapper doesn't expose (e.g. `ANY($1::uuid[])` array params, raw `NpgsqlCommand` for performance fixtures), the test owns the Npgsql type-conversion contract that Dapper used to handle silently — `DateTime.Kind=Utc` must be converted to `Unspecified` before binding into a `timestamp without time zone` column (cycle 6: AZ-505 introduced two Dapper-bypassing paths and all three new test files hit the same `Cannot write DateTime with Kind=UTC` error until `DateTime.SpecifyKind(..., Unspecified)` was added at the bind sites).
+  Source: _docs/06_metrics/retro_2026-05-12_cycle6.md
+- [2026-05-12] [testing] Tests that assert specific schema artifact names (`idx_<name>` / `pk_<name>` / `fk_<name>`) need cross-migration awareness — phrase assertions at the capability abstraction level ("any index whose first column is `location_hash`") rather than the artifact-name level when possible, otherwise drop/rename migrations require fixture co-updates in the same PR (cycle 6: `MigrationTests.Az503NewUniqueIndexCoversIntegerKeyAndFlightId` hardcoded `idx_tiles_location_hash` from migration 014; migration 015 dropped it, broke the assertion until broadened to accept either index name).
+  Source: _docs/06_metrics/retro_2026-05-12_cycle6.md
 - [2026-05-12] [architecture] Cross-repo cryptographic invariants (UUID namespaces, deterministic-key formulas, base32/64 alphabets, tile-zoom conventions) MUST live as code-level constants in BOTH repos with reference-vector tests on BOTH sides — documentation alone is insufficient because constant drift surfaces only as 100% lookup misses in production, harder to detect than a unit-test failure (cycle 5: AZ-503 introduced `TileNamespace = 5b8d0c2e-7f1a-4d3b-9c5e-1f3a8e7d2b6c` which must byte-match the same constant in `gps-denied-onboard/components/c6_tile_cache/_uuid.py`; the satellite-provider side has the constant + 10 Python-generated reference vectors in `Uuidv5Tests.cs` and the sibling repo will mirror).
   Source: _docs/06_metrics/retro_2026-05-12_cycle5.md
 - [2026-05-12] [tooling] Local Docker/colima DNS cold-start is a recurring class of failure that contaminates the Step-15 perf gate — when the perf-mode "one re-run" rule fires twice across consecutive cycles with the same root-cause class (DNS / NTP / resolver), the harness must escalate from "re-run" to a deterministic fix at the harness layer (DNS pre-warm in script, OR move gate to CI), not just another re-run (cycle 5: PT-01 failed Run #1 on `tile.googleapis.com` cold-start, then Run #2 on `mt0.google.com` cold-start; the warmup probe between runs only touched the hostnames it explicitly named).
@@ -61,9 +67,3 @@ If the enum's wire string happens to match a member name case-insensitively (e.g
   Source: _docs/06_metrics/retro_2026-05-11_cycle2.md
 - [2026-05-11] [testing] Integration tests must explicitly reset DB state at startup — relying on wallclock seeds or "tests probably won't collide" is a workaround, not isolation; the persistent Postgres volume in docker-compose makes test data accumulation the default state (cycle 2: `UavUploadTests._coordinateCounter` collision was patched with a wallclock seed instead of a real DB-reset hook).
   Source: _docs/06_metrics/retro_2026-05-11_cycle2.md
-- [2026-05-11] [testing] Persisted enums need a Dapper read-roundtrip integration test — unit-testing the type handler in isolation does not prove read-side behavior (see L-001).
-  Source: _docs/06_metrics/retro_2026-05-11.md
-- [2026-05-11] [process] NFR test-spec additions must include the runner-script implementation in the same step, or be tagged "Deferred — harness work tracked in <ticket>"; otherwise scenarios accumulate as Unverified across cycles.
-  Source: _docs/06_metrics/retro_2026-05-11.md
-- [2026-05-11] [estimation] Task-spec test-site-count estimates must be backed by an explicit grep evidence block, not pattern-matched against neighboring code (AZ-484 spec said ~3 sites in `RegionServiceTests`; actual = 0).
-  Source: _docs/06_metrics/retro_2026-05-11.md