docs: introduce ROAD-TO-V2-Overhaul documentation, update README to link it, and remove a redundant variable capture in eval_prompt_test.go.

Author: lemon07r

README.md

@@ -23,6 +23,7 @@ A lightweight evaluation harness for coding agents that runs high-signal, compac
 - [How It Works](#how-it-works)
 - [Output](#output)
 - [Architecture](#architecture)
+- [Version History](#version-history)
 - [Contributing](#contributing)
 - [License](#license)
 
@@ -280,6 +281,10 @@ sanityharness/
 
 See [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) for architecture details.
 
+## Version History
+
+For a full summary of all changes since `v1.6.1` (the entire `1.7.x` line and further), see [docs/ROAD-TO-V2-Overhaul.md](docs/ROAD-TO-V2-Overhaul.md).
+
 ## Contributing
 
 Contributions are welcome! Please see [docs/CONTRIBUTING.md](docs/CONTRIBUTING.md) for guidelines.

docs/ROAD-TO-V2-Overhaul.md

@@ -0,0 +1,79 @@
+# Road to v2.x.x Overhaul: What's Changed Since v1.6.1
+
+If you want the short version: `1.7.x` is a fairness and reliability release. We reduced information leakage during eval, made infrastructure failures easier to reason about, and tightened task contracts where hidden requirements were too implicit.
+
+## Why we changed anything
+
+By the time we reached `v1.6.1`, we had enough evaluation history to see recurring issues:
+- some outcomes were influenced by workspace visibility rather than pure task solving,
+- transient agent or infrastructure problems were being mixed with normal failures,
+- and a few tasks required behavior that was technically tested but not clearly stated.
+
+`1.7.x` is the pass where we cleaned those up.
+
+## What changed in evaluation behavior
+
+The biggest shift is workspace isolation during `sanity eval`.
+
+Agents now run in isolated temporary workspaces under `/tmp`, and then the harness copies the resulting code back into `eval-results` for validation. In practical terms, agents cannot inspect sibling tasks, prior eval outputs, or their own running log stream while solving.
+
+This is the key fairness change in `v1.7.0`. It removes a class of accidental side-channel advantages and makes comparisons more defensible.
+
+We also moved `agent.log` placement so it lives in task output directories (`eval-results/.../<task>/agent.log`) rather than inside the active workspace.
+
+## What changed in failure handling
+
+Another major improvement is how we handle infra-style failures.
+
+In older behavior, empty or broken agent runs could look like regular task failures. In `1.7.x` we detect these cases explicitly, retry appropriately, and keep resume flow clear. The end result is that pass/fail numbers better reflect model behavior, not random execution flakiness.
+
+Resume messaging and bookkeeping were also improved so interrupted or infra-affected runs are easier to continue safely.
+
+## What changed in prompts and test stability
+
+We updated eval prompts to include explicit toolchain versions. That was a practical fix for frequent version mismatch mistakes, especially in ecosystems with fast API churn.
+
+We also removed machine-dependent wall-clock assertions from hidden tests in `rust/regex-lite`. The task is still challenging, but now runtime pressure is enforced by harness/container timeouts instead of host-specific timing thresholds.
+
+## What changed in task specs
+
+Several low-pass tasks were failing for reasons that looked more like under-specified requirements than true capability gaps.  
+For those tasks, we tightened textual contracts in stubs and comments so hidden expectations are inferable without giving away implementation details.
+
+Tasks updated in this pass:
+- `tasks/typescript/promise-pool/promise_pool.ts`
+- `tasks/dart/future-pool/lib/future_pool.dart`
+- `tasks/typescript/csv-lite/csv.ts`
+- `tasks/rust/macros/lib.rs`
+- `tasks/kotlin/channel-multiplexer/src/main/kotlin/ChannelMultiplexer.kt`
+- `tasks/dart/reactive-cache/lib/reactive_cache.dart`
+- `tasks/dart/isolate-pool/lib/isolate_pool.dart`
+- `tasks/kotlin/flow-processor/src/main/kotlin/FlowProcessor.kt`
+- `tasks/zig/arena-allocator/arena.zig`
+- `tasks/zig/comptime-json/json.zig`
+
+The intent here was high-signal evaluation: remove "mind-reading" requirements, but do not turn tasks into copy-paste exercises.
+
+## Compatibility and comparing old runs
+
+`1.7.x` is intentionally not identical to `v1.6.1` behavior. If you are comparing against historical leaderboard-era runs, use legacy mode:
+
+```bash
+./sanity eval --legacy
+```
+
+Use default mode for current evaluations. Use legacy mode only when you need apples-to-apples historical comparison. 
+
+## Commit range
+
+This document covers:
+- baseline: `v1.6.1`
+- through: current `HEAD` on `main`
+
+Main commits in range:
+- `a3a2758`
+- `c69c19e`
+- `5e972ea`
+- `f192caf`
+- `3567905`
+- `ba25c9a`

internal/cli/eval_prompt_test.go

@@ -112,7 +112,6 @@ func TestBuildAgentPromptIncludesToolchainInfo(t *testing.T) {
 	}
 
 	for _, tc := range tests {
-		tc := tc
 		t.Run(tc.name, func(t *testing.T) {
 			t.Parallel()