fix(sse): stop re-prepending Kiro tool docs onto every subsequent turn (#13652)

[diegosouzapw]

Closes #13652

Root cause (short)

convertMessages() in open-sse/translator/request/openai-to-kiro.ts is stateless per HTTP
request. OpenAI-compatible clients resend the full growing message array on every turn, so the
same tool-bearing first user message gets re-scanned on every request and its relocated
documentation (toolDocs) rebuilt from scratch. buildKiroPayload() then unconditionally
prepended it onto whatever the current turn was, so the ~10KB+ doc block kept landing on the
newest user message on every subsequent turn instead of staying anchored to the turn that
originally carried it.

Fix

Track the actual turn object that ends up carrying the relocated docs (toolDocsCarrier), set at
both attachment sites in convertMessages(). After currentMessage is finalized, if the
docs-bearing turn was not promoted to currentMessage (i.e. it was demoted into history on
a later turn), embed the doc text directly onto that history entry's own content and clear the
local toolDocs so buildKiroPayload()'s existing prepend does not also inject it. When the
docs-bearing turn IS currentMessage (single-turn conversations, or the "no user turn" fallback),
behavior is unchanged. Net effect: the doc block still reaches the model on every request (no
regression of the earlier _toolDocs bug that dropped documentation entirely), but now exactly
once per request, anchored to the turn it was originally attached to instead of re-glued onto the
newest turn.

This does not reduce the total bytes sent to Kiro per request (same total payload size) — it only
fixes where the doc block lands. Cutting the actual byte/token cost would require confirming
whether Kiro's conversationId-scoped Builder ID cache retains prior-turn content server-side,
which is unverified from the repo alone and out of scope here (flagged as a follow-up).

Regression test

tests/unit/issue-13652-kiro-tooldocs-repeat.test.ts

RED (unfixed code):

✖ issue #13652: relocated tool doc is re-prepended on every subsequent turn, not just the first
  AssertionError [ERR_ASSERTION]: turn 2: the tool documentation must NOT be re-prepended to a
  later turn's content once it has already been delivered earlier in the conversation
  (found 12113 chars, incl. the doc block, in currentMessage.content: # Tool Documentation ...)

GREEN (fixed code):

✔ issue #13652: relocated tool doc is not re-prepended to a later turn once already delivered
ℹ tests 1
ℹ pass 1
ℹ fail 0

Gates run

• npx eslint --suppressions-location config/quality/eslint-suppressions.json open-sse/translator/request/openai-to-kiro.ts tests/unit/kiro-long-tool-description-docs.test.ts tests/unit/issue-13652-kiro-tooldocs-repeat.test.ts → clean, no output
• npm run check:open-sse-typecheck → openSseTypecheckErrors=0
• node scripts/check/check-file-size.mjs → no violation on touched files (one pre-existing, unrelated frozen-file violation in open-sse/utils/stream.ts, not touched by this PR)
• node scripts/check/check-complexity.mjs → OK — 2824 violations (baseline 3218)
• node scripts/check/check-cognitive-complexity.mjs → OK — 1276 violations (baseline 1437)
• node scripts/check/check-test-discovery.mjs → OK, new test file discovered
• DATA_DIR=$(mktemp -d) node --import tsx/esm --test --test-force-exit tests/unit/issue-13652-kiro-tooldocs-repeat.test.ts → 1/1 pass
• DATA_DIR=$(mktemp -d) node --import tsx/esm --test --test-force-exit tests/unit/kiro-long-tool-description-docs.test.ts → 6/6 pass
• DATA_DIR=$(mktemp -d) node --import tsx/esm --test --test-force-exit tests/unit/executor-kiro.test.ts → 20/20 pass
• Full sweep of every other test file that imports openai-to-kiro/buildKiroPayload/convertMessages (kiro-interleaved-tool-results-8903, kiro-model-aliases, kiro-system-reminder-2306, openai-to-kiro-helpers-split, translator-openai-to-kiro, openai-responses-reasoning-effort, repro-6576-kiro-thinking-unsupported-model, request-dedup-10249, translator-ai-sdk-image-parts, kiro-continue-filler-5231) → 80/80 pass

Existing tests aligned

tests/unit/kiro-long-tool-description-docs.test.ts — two tests updated:

• "relocated tool documentation reaches the current turn for every turn shape" → renamed to
  "relocated tool documentation reaches exactly one turn for every turn shape". It previously
  asserted the doc block always lands on currentMessage.content for every TURN_SHAPES entry —
  that encoded the pre-fix bug for the multi-turn shapes (where the tool-bearing turn is demoted
  into history). Now it asserts the doc block reaches exactly one turn across
  history + currentMessage combined, which is the actual regression guard for fix(providers): Kiro translator re-prepends full relocated tool documentation on every turn #13652 (catches
  both silent dropping and re-injection).
• "only oversized descriptions are relocated in a mixed tool inventory" → the assertions that
  checked current.content for the relocated doc section now check the combined
  history + currentMessage content, since the "multi-turn conversation" shape it uses now
  anchors the doc in history[0].

Both changes are alignments to the corrected contract, not weakenings — no assertion was removed
or loosened; the "exactly once" check is strictly more specific than the prior "current turn
contains it" check.