[yusiwen]

What changed

• eval/spike-repo-index/ (new, a nested module): six tasks adapted from the tier-2/3 suite, each padded
  with ten unrelated internal/ packages so a defect has to be found in a tree, and each with its tests
  in a sibling hidden/ directory instead of the work copy.
• eval/runner: -hidden <dir> installs a task's suite after the run (the work copy carries no test
  at all); a map configuration prepends a bounded symbol digest to the prompt; copyTask now carries
  subpackages, without which the padding never reached the agent.
• eval/runner/spike.go: the loader for tasks whose tests are outside them, the hidden-suite installer,
  promptFor, and repoMapDigest (files, packages, imports, exported signatures; 2000-token cap;
  deterministic).
• eval/runner/spike_test.go: proves the corpus offline — red with the hidden suite, green after the
  reference solution; padding reaches the work copy; prompts name no file; the digest's shape, budget and
  determinism; the map arm's prompt carries it and the baseline's is untouched.
• docs/repo-index-spike.md: the measurement, its verdict, the qualifications, the conditions for
  reopening, and how to reproduce it.
• docs/roadmap.md: B4 back to Not started — spiked 2026-10-28, no-go ([agent] Spike: is a repository index worth building? (roadmap B4) #308), with the verdict and
  the three reopening conditions in its entry (In progress 3 → 2, Not started 20 → 21, total still 40).
• docs/eval.md points at the corpus; the documented test count follows the tree (1270).

Why

The harness grades with a task's visible test, and a failing test prints the package and line — it
localises the defect before the agent starts. That measures comprehension, not discovery, and discovery is
the half of a repo map B4 is about. Hidden tests fix the instrument; padding gives discovery something to
cost.

Result

	build	map
Solved	6/6 (12/12 runs)	6/6 (12/12 runs)
Median tokens per run	115,649	86,978
Total tokens	1,792,029	2,416,678
Tokens per solved task	298,671	402,779

The pre-registered condition (solve ≥2 more, or cut tokens per solved task ≥20%) is not met →
no-go for an injected map. The document records why that verdict is narrower than "a repo map is
useless": the corpus had no headroom left, the total rests on one 35-step attempt while the median run
got cheaper and shorter, and the callable tool surface was never built.

Verification

• go test ./... -count=1 → 16 packages ok; go vet and gofmt -l clean.
• The corpus is proved without a provider: TestSpikeTasksAreSolvable runs every task red→green against
  its reference solution (12 go test invocations), TestInstallHiddenTestsRefusesAnEmptySuite blocks a
  vacuous pass, TestSpikePromptsDoNotNameFiles blocks a leaky prompt.
• git log --format='%h %G? %s' → G.

Honest scope

• Six tasks of twenty to thirty files, one language, one model, two repetitions per cell. The mean is not
  a robust statistic at that size, which is why the median is reported beside it.
• The tool-surface arm was not built; the issue's method named three configurations and this ran two. The
  narrowing is recorded in the document, not discovered later.
• The run needed the binary copied to /tmp (executing bin/tinycode in place fails every provider call
  on this machine, byte-identical copy succeeds); that is a harness quirk outside this repository and is
  noted in the document's reproduction section.

Follow-ups

• [agent] Spike: is a repository index worth building? (roadmap B4) #308 closes with this document as its answer; B4's entry carries the reopening conditions.
• B5's S2/S3 ([agent] Give context management turn-level control: pin, unpin, drop #306) remain open.

Refs #308, #168