Add an indicator CLI and a shared indicator registry, reused by MCP

[cinar]

Closes #415.

Summary

• New registry package (root module): a name -> indicator lookup. Given an indicator name, string parameters (e.g. Period=14, or nested paths like Ema1.Period for MACD), and a channel of asset.Snapshot, it computes the indicator and returns its named output series aligned with dates. 76 indicators registered across trend, momentum, volatility, and volume (-list on the CLI, or registry.Names(), enumerates them) — split into definitions_trend.go/definitions_momentum.go/definitions_volatility.go/definitions_volume.go for readability.
• New cmd/indicator-cli, matching the shape of indicator-sync (-source-name/-source-config for the asset repository, -last for a lookback window). Looks up an indicator by name via the registry and writes the result as CSV to stdout (dates/floats formatted the same way helper.Csv formats them elsewhere in this project); -list prints the registered names.
• New indicator tool on the MCP server (mcp/indicator.go, mcp/mcp_server.go), built on the same registry as the CLI, so both front ends stay in sync as indicators are added.
• mcp/go.mod now has a replace github.com/cinar/indicator/v2 => ../ directive, so mcp/ builds against the local tree instead of trailing the last tagged release — otherwise the new registry package would be invisible to it.
• Added ./registry/... to taskfile.yml's INDICATOR_BASE and an indicator-cli build to build-tools; added /indicator-cli and /mcp/mcp to .gitignore (root-anchored — an earlier unanchored indicator-cli entry accidentally also matched the cmd/indicator-cli/ source directory).
• Merged in master's subsequent fixes for trend.Stc: ComputeWithContext output channel never closes #417-420 (landed as Fix Stc.ComputeWithContext output channel never closing #421-424) and registered the two that turned out fully correct: momentum.Fisher (fisher) and momentum.PringsSpecialK (prings_special_k).

Not registered, on purpose

• MovingMax, MovingMin, MovingSum, MovingStd, trend.Stochastic (the generic single-series one), momentum.Streak: internal building blocks other indicators compose, not traded on their own.
• Envelope, Mlr, Mls: no natural default to construct with (a caller-supplied Ma implementation, or an arbitrary (x, y) series with no fixed OHLCV field to bind to).
• PivotPoint: returns a struct per value instead of a plain number, which doesn't fit the Series shape every other indicator here uses.
• trend.Stc, trend.T3: exercising every registered indicator against real and synthetic data surfaced that Fix Stc.ComputeWithContext output channel never closing #421/Fix T3 dropping most of its output #423 only fixed each one's channel-closing/output-length symptom, not its underlying formula. Stc divides by a denominator that's frequently near zero, yielding -Inf and swings like 1435/-710 for an indicator documented to oscillate 0-100. T3's weighted-sum coefficients don't sum to 1, so it doesn't track price at all (a flat 100-input returns -63). Filed as trend.Stc: STC formula divides by a near-zero denominator, producing -Inf and huge out-of-range swings #425/trend.T3: weighted-sum coefficients don't sum to 1, so T3 doesn't track price at all #426, with fixes up as PR Fix STC dividing by a near-zero denominator, producing -Inf/huge out-of-range swings #428/Fix T3's weighted-sum coefficients not summing to 1 #429 (not yet merged) — along with a third bug found in the process, trend.MovingSum permanently going NaN after a single bad value instead of recovering once it leaves the window (trend.MovingSum: a single NaN/Inf permanently poisons the running sum, even after it leaves the window #427, PR Fix MovingSum's running sum staying NaN/Inf forever after one bad value #430).

Design notes from review

• registry.Run feeds each indicator's input from its own independent producer (materializing the snapshots once, then one helper.SliceToChanWithContext per consumer) rather than a single synchronized fan-out (helper.DuplicateWithContext) shared between the dates and every Field. This was required to register momentum.ConnorsRsi/momentum.Rvi, which read their entire input before returning anything — sharing a lock-step fan-out with those deadlocks, since nothing drains the dates channel until Run itself returns. As a side effect, Result's Dates and Series channels are now safe to drain in any order/pace, not just interleaved.
• indicator-cli's CSV writer formats dates and floats the same way helper.Csv does (helper.DefaultDateTimeFormat, 'g' verb) instead of a hardcoded layout and 'f', so its output reads like any other CSV this project produces; writeCsv takes io.Writer instead of *os.File.
• mcp/indicator.go's runIndicator derives a cancelable context and its snapshot-producing goroutine selects on ctx.Done(). Previously, an early return (e.g. a series ending before the dates did) left the rest of the computation pipeline's goroutines blocked forever in this long-lived server process — a leak per failed request.
• registry/registry_test.go's TestRunEveryRegisteredIndicator synthetic fixture is now 800 points (was 400): prings_special_k's IdlePeriod() is 724, the longest of any registered indicator, and needs the extra runway.

Test plan

• go build ./... (root module) and go build ./... (mcp module)
• go vet, staticcheck, revive, gosec clean on the new/changed packages
• go test -race -cover ./registry/... — 95%+ coverage, including a smoke test that runs every registered indicator (all 76) end to end
• go test -race -cover ./... in mcp/ — includes an in-process MCP client test calling the new indicator tool
• go test -race ./... at the repo root — full suite, no regressions from the master merge
• Manual CLI runs against asset/testdata/repository/brk-b.csv: single-output (sma), nested-param (macd -param Ema1.Period=), multi-output (ichimoku_cloud 5 series, kdj, aroon), the two newly-registered indicators (fisher, prings_special_k), -list, and the unknown-indicator error path
• task docs was not run — it produced degraded output (lost GitHub blob-link resolution across every package) in the sandbox this was built in; worth a separate docs pass before merging, consistent with how doc updates have landed as their own PRs recently (e.g. Document update. #414).

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

❌ Patch coverage is 87.95580% with 109 lines in your changes missing coverage. Please review.
✅ Project coverage is 91.83%. Comparing base (6653727) to head (74825c6).

Files with missing lines	Patch %	Lines
cmd/indicator-cli/main.go	0.00%	78 Missing ⚠️
registry/registry.go	65.16%	25 Missing and 6 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #416      +/-   ##
==========================================
- Coverage   92.32%   91.83%   -0.49%     
==========================================
  Files         229      235       +6     
  Lines        7280     8185     +905     
==========================================
+ Hits         6721     7517     +796     
- Misses        472      575     +103     
- Partials       87       93       +6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.