zfin/TODO.md

# Future Work

No work here is blocking — we're in a good state. Items below are
ordered roughly by priority within each section. Priority labels
(`HIGH` / `MEDIUM` / `LOW`) mark items that deserve explicit
ranking; unlabeled items are "someday, if the mood strikes."

**Next up:** configurable benchmark symbols (low-effort win) and the
manual-check accounts mechanism (medium effort, real user value).

## Projections: future enhancements

- **Configurable benchmark symbols — priority MEDIUM.** Currently
  hardcoded SPY + AGG. Route through `projections.srf` as a
  `type::config,benchmark::SYMBOL` record (or similar). Low effort.
- **Configurable return cap per position — priority MEDIUM.**
  Default: none; cap outliers like NVDA. Should route through
  `projections.srf` cleanly.
- **"Not Retired Yet" accumulation-phase mode — priority HIGH.**
  Accumulation phase with contributions before retirement.
  Separate from life events — the simulation has two distinct phases:
  accumulation (base spending = 0, contributions applied) and
  distribution (searched-for withdrawal + spending model). Config:
  `retirement_age::60` or `retirement_in::10`, plus
  `annual_contribution::100000`. Safe withdrawal search only applies
  to the distribution phase. Chart shows portfolio growing during
  accumulation, peaking at retirement, then drawing down.
- Configurable MIN period selection (currently 3Y/5Y/10Y, exclude 1Y)
- Multiple spending models: flat (current), decreasing (1-2% real annual decrease,
  Blanchett "spending smile"). Late-life healthcare better modeled as a life event.
- Unclassified position handling in allocation split (warn user)
- Historical projection comparison: re-run projections from any past snapshot date,
  overlay actual portfolio trajectory from subsequent snapshots onto the projected
  percentile bands. Shows how reality tracked against the model. Data is already
  available in history/*.srf snapshots — just need to load a historical portfolio
  value and re-run `computePercentileBands` with that starting point, then plot
  actual values from later snapshots as a line overlaid on the bands.

  `zfin projections --as-of <DATE>` already reruns the simulation
  against a past snapshot (the prerequisite for this overlay). What's
  missing is the overlay itself — loading multiple downstream snapshots
  and plotting their net-worth trajectory on the same chart.

  **Deferred to ~2027.** Needs a practical volume of real snapshots
  (currently building up; meaningful backtest requires 12+ months).
  Backfilling from git history is not viable — the lot-level state on
  portfolio.srf at a past commit is insufficient to reconstruct the
  full transaction+contribution picture. Revisit once there are 12+
  months of continuous snapshot data.

  Also consider: `metadata.srf` and `projections.srf` classifications /
  assumptions drift over time. For back-dated runs we currently use
  the current versions of both; historical git-tracked versions could
  be checked out and loaded instead. Edge case for now.

## Audit: manual-check accounts mechanism (NYL, Kelly's ESPP, etc.) — priority HIGH

Some accounts/positions can't be reconciled from broker CSVs and need a
human-in-the-loop reminder at the audit step. Examples:

- **NY Life** — no CSV export at all. Values only live in periodic
  statements.
- **Kelly's ESPP** — accrued payroll-deduction cash doesn't appear in the
  Fidelity positions CSV until the purchase date hits (typically every
  6 months). Between purchases the cash is a real contribution that
  `zfin audit` can't see.
- Future: treat as an open category.

The existing `update_cadence::weekly|monthly|quarterly|none` field already
sort-of covers this, but has two gaps:

1. It fires off the last *git-detected change*, not the last *human
   review*. For NYL, the value sometimes hasn't changed in months — so
   git never fires, cadence never trips.
2. ESPP needs weekly-ish attention while accumulating cash between
   purchases, but the accrued balance is invisible to the CSV audit.

### Options

A. **New `update_cadence::manual` variant** — always fires every audit
   run until silenced. Blunt but zero design work.

B. **`last_refreshed::YYYY-MM-DD` field on `accounts.srf`** — explicit
   human-review timestamp, decoupled from git-detected changes. Audit
   compares `today - last_refreshed` against the cadence. User bumps
   the field when they check the statement. Probably the most
   correct fit for NYL.

C. **Sticky TODO list** — a `todos.srf` or `todo::` field on accounts
   that audit always surfaces until cleared. General-purpose; also
   covers "remember to rebalance on 5/15".

### ESPP-specific follow-through

ESPP is also a contribution-attribution blind spot. If Kelly's paystub
deducts $X/week but the cash lot doesn't reach `portfolio.srf` until
the purchase date, the attribution math is under-counting contributions
and over-counting the purchase-week gain. Possible fixes are discussed
in the "Contributions diff" TODO below — option C there (per-account
`cash_is_contribution`) would make manually-entered ESPP
cash additions count correctly.

## In-kind transfer support (`type::in_kind`) — priority MEDIUM

`transaction_log.srf` parses `type::in_kind` records but the
contributions matcher always rejects them with "in-kind transfers
not yet supported in v1." In-kind movements need per-symbol
matching across accounts: an in-kind transfer of 100 VTI shares
from Acct A to Acct B shows up as `lot_removed` on A + `new_stock`
on B (or a `rollup_delta` share increase if B already had a VTI
lot), neither of which can be matched by the current
amount-based cash matcher.

Proposed: a second pass in `matchTransfers` that iterates
`type::in_kind` records and looks for same-symbol matches across
`lot_removed` on `from` + `new_stock`/`rollup_delta` on `to`
within the window. Gated on share-count and open_price sanity so
a partial transfer doesn't false-positive against an unrelated
edit.

Driver: when the user starts moving positions between accounts
directly (e.g. Roth conversion of already-held shares, 401k →
rollover IRA in-kind) rather than liquidating and re-buying.

## Torn SRF files from server sync (root cause unknown)

**Status:** Root cause still unidentified. We have mitigations and
diagnostics in place that keep torn responses from corrupting the
cache, but we don't yet know *why* responses arrive torn. Until we
have a root cause, this is not resolved — it's mitigated.

Mitigations landed so far:

- `syncFromServer` (`src/service.zig`) validates responses via
  `cache.Store.looksCompleteSrf` before `writeRaw`. Torn HTTP bodies
  (empty, missing `#!srfv1` header, or no trailing newline) are
  rejected with a warn-level log and NOT written to cache.
- HTTP responses are checked for an `ETag` sha256 header; on mismatch
  we retry the request once before giving up and falling back to the
  provider.
- Read-path self-heal: on SRF parse failure during read, the cache
  entry is invalidated so a subsequent refresh can repair without
  user intervention.
- Diagnostics: richer error capture around the sync path. So far,
  HTTP transit is the dominant source of torn responses — but that's
  an observation, not a root cause.

**Remaining work:**

- Identify root cause. Candidates to investigate: proxy/load-balancer
  behavior, HTTP keepalive reuse, partial reads on the server side,
  client-side buffer handling. The etag retry tells us whether the
  problem is per-request or persistent; dig into the diagnostics
  output when the next occurrence is captured.
- Once root cause is known, decide whether the current mitigations
  are sufficient or whether a targeted fix is needed. The
  mitigations may end up being the whole answer, but we can't
  conclude that without understanding the underlying cause.

(Content-Length validation was considered and rejected: once the
server starts compressing response bodies, Content-Length reflects
the compressed byte count, not the decoded payload, so it's not a
reliable integrity check.)

## Market-aware cache TTL for daily candles

Daily candle TTL is currently 23h45m, but candle data only becomes meaningful
after the market close. Investigate keying the cache freshness to ~4:30 PM
Eastern rather than a rolling window. This would avoid unnecessary refetches
during the trading day and ensure a fetch shortly after close gets fresh data.
Probably alleviated by the cron job approach.

## On-demand server-side fetch for new symbols

Currently the server's SRF endpoints (`/candles`, `/dividends`, etc.) are pure
cache reads — they 404 if the data isn't already on disk. New symbols only get
populated when added to the portfolio and picked up by the next cron refresh.

Consider: on a cache miss, instead of blocking the HTTP response with a
multi-second provider fetch, kick off an async background fetch (or just
auto-add the symbol to the portfolio) and return 404 as usual. The next
request — or the next cron run — would then have the data. This gives
"instant-ish gratification" for new symbols without the downsides of
synchronous fetch-on-miss (latency, rate limit contention, unbounded cache
growth from arbitrary tickers).

Note that this process doesn't do anything to eliminate all the API keys
that are necessary for a fully functioning system. A more aggressive view
would be to treat ZFIN_SERVER as a 100% source of record, but that would
introduce some opacity to the process as we wait for candles (for example) to
populate. This could be solved on the server by spawning a thread to fetch the
data, then returning 202 Accepted, which could then be polled client side. Maybe
this is a better long term approach?

## Low-priority items

The following items are acknowledged but not prioritized. Listed here
so they don't get lost; pick up opportunistically.

### UX

- **CLI options command UX.** The `options` command auto-expands only
  the nearest monthly expiration and lists others collapsed. Reconsider
  the interaction model — e.g. allow specifying an expiration date,
  showing all monthlies expanded by default, or filtering by strategy
  (covered calls, spreads).
- **TUI: toggle to last symbol keybind.** A single-key toggle that
  flips between the current symbol and the previously selected one
  (like `cd -` in bash or `Ctrl+^` in vim). Store `last_symbol` on
  `App`; on symbol change, stash the previous. Useful for
  eyeball-comparing performance/risk data between two symbols.

### Data quality

- **Fix `enrich` command for international funds.** `deriveMetadata`
  in `src/commands/enrich.zig` misclassifies international ETFs:
    1. `geo` uses Alpha Vantage's `Country` field, which is the *fund
       issuer's* domicile (USA for all US-listed ETFs), not the fund's
       investment geography. Every US-domiciled international fund gets
       `geo::US`.
    2. `asset_class` short-circuits to `"ETF"` when
       `asset_type == "ETF"`, or falls through to a US-market-cap
       heuristic that always produces `"US Large Cap"` /
       `"US Mid Cap"` / `"US Small Cap"`.

  Known misclassified tickers (all came back as
  `geo::US, asset_class::US Large Cap`):
    - **FRDM** — Freedom 100 Emerging Markets ETF → should be
      `geo::Emerging Markets, asset_class::Emerging Markets`
    - **HFXI** — NYLI FTSE International Equity Currency Neutral ETF
      → should be
      `geo::International Developed, asset_class::International Developed`
    - **IDMO** — Invesco S&P International Developed Momentum ETF →
      should be
      `geo::International Developed, asset_class::International Developed`
    - **IVLU** — iShares MSCI International Developed Value Factor
      ETF → should be
      `geo::International Developed, asset_class::International Developed`

  The Alpha Vantage OVERVIEW endpoint doesn't provide fund geography
  data. Options: use the ETF_PROFILE holdings/country data to infer
  geography, parse the fund name for keywords ("International",
  "Emerging", "ex-US"), or accept that `enrich` is a scaffold and
  emit a `# TODO` comment for ETFs instead of silently misclassifying.

### Options / valuation

- **Per-account covered call adjustment.** `adjustForCoveredCalls` in
  `valuation.zig` operates on portfolio-wide aggregated allocations.
  It matches sold calls against total underlying shares across all
  accounts. This is wrong — calls in one account can only cover
  shares in that same account. Fixing means restructuring
  `portfolioSummary`, since `Allocation` is currently
  account-agnostic. Low priority — naked calls are rare, and calls
  are typically in the same account as the underlying.
- **Covered call adjustment O(N*M) loop.** `adjustForCoveredCalls`
  has a nested loop — for each allocation, it iterates all lots to
  find matching option contracts. Fine for personal portfolios
  (<1000 lots). Pre-indexing options by underlying would help if
  someone had a very large options-heavy portfolio.

### Analysis / correctness

- **Analysis account/asset-class total mismatch.** The "By Account"
  and "By Tax Type" sections in the analysis command sum to slightly
  more than "Asset Class" (~0.6% error). Likely a discrepancy between
  how the lot-level account loop values cash, CDs, or options vs how
  the asset-class section computes them via `portfolio.totalCash()` /
  `totalCdFaceValue()`. Per-account values themselves are correct
  after the price_ratio fix.

### Audit

- **Audit large-lot threshold tuning.** `src/commands/audit.zig` uses
  `audit_large_lot_threshold: f64 = 10_000.0` as the cutoff for
  "surface this new lot for confirmation." Revisit if $10k proves too
  aggressive (ESPP accruals spam the report) or too permissive (large
  DRIP confirmations slip past). If runtime tuning becomes necessary,
  a `--large-lot <amount>` flag or a global
  `audit_large_lot_threshold` field on `accounts.srf` would be
  reasonable extensions.

### Infra / performance

- **HTTP connection pooling.** Parallel server sync in `loadAllPrices`
  spawns up to 8 threads, each with its own HTTP connection. Could
  reuse connections to reduce TCP handshake overhead. Only matters
  with very large portfolios (100+ symbols) hitting ZFIN_SERVER.
- **Streaming cache deserialization.** Cache store reads entire files
  into memory (`readFileAlloc` with 50MB limit). For portfolios with
  10+ years of daily candles, this could use significant memory.
  Keep current approach unless memory becomes a real problem.