refactor(prompt): make prompt construction template-driven

[mariusvniekerk]

Why

Prompt construction in internal/prompt had accumulated too much behavior in Go-side string assembly, which made the actual prompt structure hard to reason about, easy to regress during cleanup, and split across multiple sources of truth. This refactor moves the prompt format itself into templates so the rendered prompt shape is defined where it is authored, while preserving existing review behavior and fixing the regressions uncovered during the migration and subsequent roborev reviews.

Summary

• make prompt construction template-driven end to end by moving assembled prompt bodies, prompt sections, and agent-specific system prompts onto embedded .md.gotmpl templates backed by typed view data
• preserve existing prompt behavior while fixing cleanup regressions in dirty and range prompt fitting, including truncated diff fallbacks, optional-context retention, and fallback selection quality
• resolve the accumulated roborev findings on gotmpl so the branch ends with no open review jobs

Test Plan

• go fmt ./...
• go vet ./...
• go test ./...

[roborev-ci]

roborev: Combined Review (9e36436)

Verdict: No medium-or-higher issues found across the submitted reviews.

The reviewed changes appear clean. Both substantive review outputs reported no issues, and the security review explicitly found no externally exploitable concerns in the modified paths under this repo’s local-only trust model.

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (04449e4)

Summary verdict: Changes are mostly sound, but there is one medium-severity regression in range prompt trimming that should be fixed before merge.

Medium

• internal/prompt/prompt_body_templates.go:287
trimOptionalSections no longer considers InRangeReviews, even though buildRangePrompt now stores per-commit reviews in optional.InRangeReviews. When a range prompt exceeds budget, the fitter can now strip subjects and current range entries before dropping this optional context, which regresses prior behavior where the optional block was trimmed first.
  Fix: include InRangeReviews in trimOptionalSections, trim it before mutating Current range metadata, and add a regression test for oversized range prompts with populated per-commit reviews.

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (21a38f8)

Verdict: No Medium, High, or Critical issues found.

All review outputs are clean. I did not find any deduplicated findings at Medium severity or above to include in this PR comment.

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (4bd884c)

No medium-or-higher issues found; the reviewed changes look clean.

All provided reviews reported no medium, high, or critical findings.

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (488ff49)

Verdict: No medium-or-higher findings; the reviewed changes look clean.

No Medium, High, or Critical issues were identified across the submitted review outputs.

---

Synthesized from 3 reviews (agents: codex, gemini | types: default, security)