Model selection strategy blog post (KR+EN) + Critic model attribution fix

Goal

• Instead of a deferred blog topic, adopt the user's real-time retrospective to write a "model selection strategy" blog post in KR+EN.
• Thesis: after switching to Fable 5 (mythos-based) on launch day and running all of S238~243 on Fable 5, results were good but a token shortage was felt → "keep daily sprints on Opus/Sonnet, and over the mid/long term inject a top-performance premium model as a targeted deployment for periodic large refactoring, which is better for long-term maintenance cost" (paying down technical-debt principal).
• Apply the blog pattern established in Sprint 237 (tone register alignment + crosscheck --strict + blog Critic).

Background

• At the /start stage, one of three deferred topics (CS quiz S215229 / things-built-then-deleted S185189·193·201 / zstd) was to be chosen, but the user paused the plan and shared their Fable 5 experience directly: "I used the newly-released Fable 5 (mythos-based) today for security and refactoring; results were good but I ran into token shortage. I'll keep using Opus/Sonnet for now, but over the mid/long term, occasionally doing a big refactor with the top-performance model might be better for long-term maintenance cost."
• This retrospective was judged to be a strong blog topic in itself — timeliness (Fable 5 launch day), real ADR grounding (S238~243 were all the Fable 5 experiment), and continuity with the existing post toward-model-agnostic-harness (order 9).
• Confirmed direction via AskUserQuestion → "replace with the model strategy retrospective." The deferred topic "things-built-then-deleted" is postponed to the next sprint.

Decision

D1. Topic swap — model selection strategy retrospective (user)

• Adopted the user's retrospective over the deferred topic. Rationale: solid ADR grounding (the S238 "switched to fable model" transition record is real) + timeliness.

D2. category·meta — ai-agent / order 11 / 3 relatedAdrs

• category ai-agent (same as toward-model-agnostic-harness, a model/agent topic).
• order 11 (prior max 10 = sliding-window). relatedAdrs ["sprint-238","sprint-241","sprint-242"] — Fable transition+audit, BE decomposition, FE decomposition respectively.

D3. Strategy insight memory SSOT separation

• The model operations strategy is recorded as SSOT in memory/project-model-selection-strategy.md separately from the blog post (by Oracle directly). A reusable reference asset.

D4. (Post-merge) Critic model attribution fix — Critic is Codex (user-caught)

• After merge, the user caught that body L48 "the wider context Critic sees at once" mis-attributes Fable 5's token consumption to the Codex-based Critic. Per CLAUDE.md, Critic is a Codex(gpt-5)-based secondary review, a separate model from Fable 5. The subject that hit the session limit was the main agent (Gatekeeper) running on Fable 5.
• Fix: replaced with "the model holds a wider context in a single pass" (EN) / "모델이 한 번에 안고 추론하는 컨텍스트" (KR). The other 4 Critic mentions (L66 redis.keys discovery · L71 cumulative rounds · L121/125 harness seat) are all accurate (the actual role of the Codex Critic) — unchanged.

Implementation

First draft (Scribe, PR #435 squash → 32f24fa)

• New: blog/content/posts/model-selection-strategy.mdx (KR, 11.3K) + blog/content/posts-en/model-selection-strategy.mdx (EN, 10.3K), 2 files +258.
• Title "Which Model, When, Where — A Strategy for Targeted Premium Deployment." Narrative: intro (S243 session-limit scene) → ADR-030 audit (High Risk 0 · Medium 3 · Low 5 · improvements 7) → 5-sprint roadmap PhaseTimeline → "regressions cluster at extraction boundaries" (S241/242) → 3 ROI conditions → balancing caveats (full refactoring's diminishing returns · cost is real) → toward-model-agnostic-harness connection.
• MDX components: PhaseTimeline/PhaseMilestone, MetricGrid/MetricCard, Callout×2, Problem/Decision/Result.
• Scribe self-review fixed 3 P1 facts: ① session-limit incident S241→S243 ② merge-gate round attribution (translation key=auto-critic · setError=R1 · derived state=R2 · R3 CLEAN) ③ S239 timeline DLQ→events API DTO.

Second fix (Oracle directly, PR #436 squash → c1d70cc)

• D4 model-attribution fix, 2 lines (KR+EN L48). Being a 2-line precise correction, Oracle handled it directly as a simple file edit (re-enacting the post's own thesis that "simple work goes on the cheap path").

Verification

• First (#435): crosscheck --strict 0 violations · blog SSG build EXIT=0 (both routes generated) · doc-refs 463 no broken · Critic (Codex gpt-5.5 --base 4c8693e) CLEAN (full fact cross-check · KR↔EN 1:1 · tone alignment) · CI 38 pass/0 fail · post-merge main run success.
• Second (#436): crosscheck --strict 0 · SSG 423 pages success · CI fail 0 · post-merge green.
• Oracle direct memory cross-check: all core facts consistent (session limit S243 · Fable transition S238 · ADR-030 numbers · S241/242 decomposition scale · cumulative rounds BE 7+FE 8=15).

Deferred

• Blog topic "things-built-then-deleted" (S185~189 ADR graph & search built → S193 graph deleted · S201 search deleted retrospective) → next sprint.
• Consider adding a model/agent attribution check to the fact-consistency gate.
• Existing backlog: (harness slot) pane guard permanence + window decoration root fix (--full FAIL 1) + Codex model pin + status misrecord 3-in-a-row · Quality — docs required promotion · sync logging singleton context · i18n · ConfirmStep tErrors · inline style (S242 carryover) · Q-4 libs/ spike (backlog) · CI helper unit test policy · (user console) GA4 3 items · live SEO redeploy · harness cron · webhook regenerate · cumulative UAT.